An Irrigated Area Map of the World (1999) derived from Remote Sensing *
Biradar, C.M., Turral, H., Noojipady, P.,
Li, Y.J., Vithanage, J., Dheeravath,
V., Velpuri, M., Schull M., Cai, X. L., Dutta, R.
A Global irrigated area map
has been produced for a nominal year of 1999 using multiple satellite sensor
and secondary data. Multiple resolution time series data used in the study
were: (a) AVHRR 4-band and NDVI 10-km monthly time series for 1981-1999, (b)
SPOT vegetation NDVI 1-km monthly time series for 1999, and (c) East Anglia
University Climate Research Unit Rainfall 50-km monthly time series for
1961-2000. Additional major global data sets used were (a) GTOPO-30 1-km
elevation, (b) JERS SAR data for the rainforests during two seasons in 1996,
and (c) University of Maryland Global Tree Cover 1-km data for 1992-93.
A number of new methods and
techniques were developed. The study first segmented the world into climate and
elevation zones and analyzed satellite images separately for these zones. The
class identification and labeling process began with spectral matching
technique (SMTs). Since time-series data are analogous to hyperspectral data, we
adopted hyperspectral analysis techniques such as SMTs to identify, group, and
label classes with similar time series characteristics. The time-series spectra
of classes were also compared with the target ones obtained from ground truthed
locations. The spectral correlation similarity was found to be the most useful
spectral matching technique (SMT). Classes are then ?verified?, at 30-50 randomly
chosen locations that are well distributed across the globe, by inspection of Google
Earth images for which the resolution varies between sub-meter to 30-meter.
Multiple image interpretation
techniques such as bispectral plots, space time spiral curves (ST-SCs),
time-series plots of normalized difference vegetation index (NDVI), and a host
of secondary data (e.g., national and global land\use and land cover data) were
used, including ESRI 150-m Landsat Geocover mosaic of the world.. Broadly
sourced ground truth data were used in identifying, labeling and refining classes.
First: IWMI?s primary ground truth data set of nearly 2000 points that include:
a) three missions conducted in 2004 and 2005 that cover the whole of India; b) extensive
data from river basins with extensive irrigation areas such as the Ganges and
Krishna in India, Ruhuna in Sri Lanka, Syr Darya in Central Asia and Limpopo in
Southern Africa; and c) a past data catalogue from the Middle East and 14
Countries in West Africa. Second: data sourced from the Degree Confluence
Project with about 4000 points that collates land use data for 1 by 1 degree
tile over the globe. In addition nearly 11,000 ?zoom in views? of high or very
high resolution Google earth points. Decision tree algorithms, NDVI time series
plots, NDVI thresholds, principal component analysis, unsupervised clustering
algorithms, and GIS spatial modeling using data such a agroecological zones,
temperature, precipitation, evapotranspiration, and elevation were widely used
to define and refine classes especially to resolve mixed classes.
A 28-class dis-aggregated
global irrigated area map at 10-kilometer scale (GIAM10km-28classes) and
aggregated 8-class and 3-class (GIAM10km-8 classes and GIAM10km-3 classes) maps
of the world were produced. The GIAM10km-28 classes (Figure 33) provide information on watering method (irrigated or rainfed agriculture), irrigation type (surface
water, ground water, and conjunctive use), irrigation intensity (single,
double, or continuous crop), and crop type or dominance. The GIAM10km-8 class
(Figure 34) provides watering method, irrigation type, and intensity. The GIAM10km-3
classes provide information on: surface water irrigation, ground water
irrigation, and conjunctive use (surface and ground water) irrigation. Informal
(e.g., small reservoirs, tanks, ground water) irrigation was identified and
mapped in addition to more conventional large scale surface water irrigation
found in most irrigated area maps. Annualized irrigated areas (intensity of
irrigation) were calculated using time-series satellite imagery from which one
can detect how many crops are grown in a same area during a given year.
Particular strengths of this work are in: (a) establishing seasonal and
annualized irrigated areas (or intensity of irrigation), (b) mapping informal (e.g., small reservoirs, tanks, ground water) irrigation in addition to conventional
surface water irrigation, (c) determining irrigated crop calendar, (d) studying
historical (e.g., last 20 years, every month) biomass dynamics for every
irrigated area class and every pixel within that class
The irrigated areas in these
maps were calculated based on sub-pixel areas (SPAs). The SPAs, which are areas
actually irrigated, were established by multiplying the full pixel areas (FPAs)
of the classes with the irrigated area fractions (IAFs) established based on:
(a) Google earth estimate (GEE), (b) high resolution imagery (HRI), and (c)
sub-pixel decomposition technique (SPDT). The combined coefficients from the
IAF-HRI and IAF-SPDT for each of the 28 GIAM classes were used to compute
robust and reliable estimates of the seasonal and annualized irrigated areas of
the world. The annualized areas are summation of areas from different seasons.
Cropping calendar (i.e., single, double, or continuous cropping) for each of
the 28 classes were established and their SPAs for each of the season were determined
by multiplying the FPA with the combined IAFs from HRI and SPDT. The annualized
area is then the sum of the areas from different seasons. For GIAM this was sum
of areas from seasons consisting of single, double, or continuous cropping. The
IAF-GEE method provides total area available for irrigation (TAAI).
The annualized irrigated areas
of the world at the end of the last millennium were 480 Mha. Of which there
was: (a) 263 million hectares (Mha) for season 1, (b) 176 Mha for season 2, and
(c) 41 Mha for continuous crops. The total area available for irrigation at any
given time at the end of last millennium was 412 million hectares of which
different proportion of areas are irrigated during different seasons as
reported above, leading to an annualized area. The distribution of irrigated
areas is highly skewed amongst continents and Countries. Asia accounts for 78 percent
(375 Mha) of all annualized irrigated areas, followed by Europe (8 percent) and
North America (7 percent). South America (3 percent), Africa (2 percent), and
Australia (2 percent) have very low proportion of the global irrigation. China has 108 Mha and India has 100 Mha of total area available for irrigation. In this 108
Mha; China has 76 Mha of crops during season 1, 68 Mha during season 2, and 7
Mha during season 3 for a annualized sum of 151 Mha. India has annualized sum
of 132 Mha, with a break up of 73 Mha during season 1, 54 Mha during season 2,
and 5 Mha during season 3. China and India have a staggering 59 percent of the
Global annualized irrigation. This is followed by USA (5 percent), Russia (3.5 percent), and Pakistan (3.3 percent). Nine other Countries (Argentina, Australia, Russia, Bangladesh, Turkey, Kazkhstan, Myanmar, Uzbekistan, and Vietnam) have areas between 1 to 2 percent. All other Countries in the
World have less than 1 percent area irrigated relative to global annualized
total. Sixty one percent of all irrigation is by surface water with the rest
(39 percent) coming from conjunctive use (surface and ground water) or ground
The accuracies of the
irrigated areas were determined using three independent datasets: (a) first, a
1861 ground truth data points of the world sourced from degree confluence
project, (b) second, a 890 point ground truth data collected by GIAM team, and
(b) third, a randomly picked 670 point ?zoom in views? of very high resolution
imagery from Google earth. From these 3 methods, the accuracies varied between
84 to 91 percent with errors of omission of not exceeding 16 percent and errors
of commission of less than 21 percent.
The IWMI Global irrigated area
map (GIAM) 28-class (GIAM10km-28 classes), 8-class (GIAM10km-8 classes), and
3-class (GIAM10km-3 classes) maps, data and products are made available through
a dedicated web portal: http://www.iwmigiam.org.
These products are supported by class characteristics (e.g., cropping
calendar), snap-shots of higher resolution imagery, time-series animations of
classes showing their biomass dynamics for last 20 years, area estimation
methods, accuracy assessment results, ground truth data links for classes
including digital photos, source images, and background documentation on
methods and materials. In addition to GIAM, data and products are also
available for: (a) a Global Map of Rainfed Cropland Areas (GMRCA), (b) a Global
Map of land use/land cover areas (GMLULCA), and (c) a generic IWMI 951-class
Global land use/land cover (LULC) Map of the World.
This document summarises the
materials and methods used to create a series of maps of irrigated areas of the
world using remote sensing approaches. These maps are complementary to existing
statistics (FAO-Aquastat) and the GIS derived maps (FAO-University of Frankfurt Global irrigated area map). The document also provides details of how the
estimates of global irrigated areas in one main season (net) and more than one
season (intensity or annualized) were derived.
The major products were a: (i) 28
class irrigated area class map (GIAM10km-28 class), comprising watering
method (in this case irrigated), irrigation type (surface water, ground water,
and conjunctive use), irrigation intensity (single, double, or continuous
crop), and crop type; (b) 8 class irrigated area class map (GIAM10km- 8
class), comprising watering method, irrigation type, and intensity; and (c) 3
class irrigated area class map (GIAM10km- 3 class) comprising surface,
ground, and conjunctive use irrigation. The estimation of seasonal global
irrigated areas is based on these products. The simpler GIAM10km-8 class and GIAM10km-3
class maps have more ?practitioner friendly? classes and are produced, to allow
The GIAM10km-28 classes, GIAM10km-8
classes, and GIAM10km-3 classes products are derived from a generic land use
and land cover (LULC) map of the world that has 951 classes and a considerable
part of the methodology is concerned with the development of this map and
subsequent definition, naming, and aggregation of those classes. The work has
had the explicit intention, as far as is possible, to take account of the
effect of cropping intensity or irrigated areas from different seasons within a
given year. Time-series analysis of remote sensing allows the basic
developmental phenology of different crops to be identified, and the number of
crop seasons in one year can be determined on aggregate for any pixel. In this
study, we have used multiple types of imagery and masking data at different
Although the analysis has been
conducted at a nominal scale of 1-km per pixel, the major source of data has
been a 20-year time series of 10-km AVHRR data. This has necessitated the use
of a classical land-use land-cover (LULC) classification approach that defines LULC
classes as a mix of land cover types. Sub-pixel disaggregation of the component
irrigation areas therefore becomes a major objective in trying to accurately
assess actual area.
The same processes and data were
used to produce the following products.
- Disaggregated 323 class Global Irrigated Area Map (GIAM10km-
- Disaggregated 229 class Global Map of Rainfed Cropped
- Aggregated 22 class map of Global Map of Rainfed Cropped
- Disaggregated 76 class Global Map of Land Use/Land Cover
- Aggregated 10 class Global Map of Land Use/Land Cover
The work has produced other significant bi-products which,
along with the main maps, are available via a dedicated website: http://www.iwmigiam.org
The website includes maps, images,
class characteristics, sub-pixel area (SPA) estimation approaches, digital
photos, ground truth data, animations of time series, and accuracy assessments.
All the background documentation is also provided.
The website contains a daunting
amount of information and data, with substantial improvements and refinements
in the presently published version 2.0. Aside from the production of the maps
and estimation of the irrigated areas, the intention of this work is to:
- provide repeatable and robust methods
and techniques of analysis of irrigated areas, and
- encourage practitioners and
researchers with better local knowledge to improve the definition and
detail in their localities and contribute to further refinement of the
This paper continues with a brief
background (section 3) to past efforts to assess irrigated areas and the
rationale for developing new approaches using remote sensing at a global scale.
In section 4 and its sub-sections, we present the basic remote sensing and other
data used to produce the maps. In section 5 and its sub-sections, we provide
details of the analytical methods applied to define and refine the classes. This
is followed by section 6 on class aggregation and section 7 on area
calculations and sub-pixel decomposition techniques (SP-DCT). Accuracies in
section 8, results and discussions in section 9 and 10, class naming convention
in section 11, products in section 12, and conclusions in section 13
3. Background and rationale
3.1 Irrigation development and trends
Following the end of the second
World War, and a period of de-colonization, there was a boom in irrigation
development which coincided with strongly motivated nation building,
particularly in Asia. Irrigated area increased at about 2.6 % per annum from a
modest 95 M ha in the early 1940s to between 250 and 280 M ha in the early
1990s (van Schilfgaarde, 1994, Siebert, S., Döll, P., Hoogeveen, J., 2002).
In this era, a key developmental
agenda for many countries was the construction of large and small dams and
river diversions to abstract and store water for agriculture. Over 40,000 large
dams (>15 meter in height) irrigate about 30-40 percent of World?s irrigated
areas (www.dams.org) and are complemented by an estimated 800,000
smaller dams. Since the 1980s, there has been a progressive decline in public
and international donor funding for irrigation, which has been replaced in many
countries by the private development of groundwater irrigation based in the
availability of cheap drilling and pumping technologies. India now has an estimated 20 million tubewell irrigators, accounting for as much as 60% of
the irrigated area according to some estimates (Shah, 2002).
This development has allowed food
production to keep pace with rapidly growing global populations and an
increasingly urban world. Farmers currently produce enough to feed the world,
although poverty and malnutrition still affect more than a fifth of the global
population, due to local shortages and inadequate distribution and market
systems. Although rates of population increase are now slowing and it is
expected that the world will continue to be able to feed itself (FAO, 2004),
there will be continued pressure to either expand irrigated area, or increase
crop and livestock productivity or substitute intensive irrigation with better
and more extensive rainfed agriculture.
The population of the World is
now approaching 6 billion and is expected to near 8 billion by 2025. To meet
future food demand, some estimate that at least another 2000 cubic kilometers
of water (equivalent to the mean annual flow of 24 additional Nile rivers) will
be needed (Postel, 1999). Water use for irrigation varies considerably across
the globe. It accounts for 2-4 % of diverted water in Canada, Germany and Poland but is an impressive 90-95 % in Iraq, Pakistan, Bangladesh, Sudan, Kyrgyzstan, and Turkmenistan (Merrett, 2002).
Globally, the irrigated landscape
remains very dynamic. Although the annual rate of increase of irrigated areas
has slowed to about 1 %, this still represents an increase of between 2 and 3
million hectares each year. There is a smaller corresponding annual loss of
irrigated area to salinity and water logging as well as abandonment of
uneconomic projects. Countries such as China and India continue to build large
multi-purpose dam projects that also supply water for irrigation. In sub-Saharan
Africa irrigation is perennially seen as having unfulfilled potential.
Elsewhere in the world there are moratoria on dam building and even the
decommissioning of dams in the western USA.
Better technology, advances in agronomy
and crop breeding (including genetically modified crops) are expected to
contribute to increasing crop land and water productivity. However, both
extensification and intensification are increasingly questioned by
environmental activists and more ecologically sensitive governments. A key
challenge for the irrigation sector lies in using less water to produce more
food, whilst mitigating negative impacts on the environment, particularly
The irrigated landscape of the
world will be shaped increasingly by the effects of competition for water from
other sectors, notably urban and rural domestic water supply and industrial
needs. It is becoming increasingly common for river basins to be
over-allocated, with negative downstream effects of competitive upstream
development, such as in the Krishna basin in India (Biggs et al, 2006).
Similarly groundwater is being mined in many places, notably significant parts
of India and in the Olgalala aquifer in the mid west of the USA. Reservation and re-allocation of flows for environmental purposes will in the end
place even greater competing demands in terms of water volume. Climate change
will impose additional challenges that will reshape the irrigated landscape
through changes in snow-melt runoff and rainfall.
In summary, irrigation is widely thought to provide 40% of
the world?s food from around 17% of the cultivated area. Key questions
concerning the sector include:
How much irrigation do we have now?
How much do we need in the future?
How much do we want in the future to achieve a sustainable
balance with the environment?
How much water does it require and will this be available?
3.2 Estimates of irrigated area
There remains considerable
uncertainty about the exact extent, area and cropping intensity of irrigation
in different parts of the world, due to the dynamics referred to above and
systematic problems of under and over-reporting of irrigation in different
contexts (e.g., ground water) and countries.
Currently, there is one irrigated
area map of the World produced by FAO/University of, Frankfurt (http://www.fao.org/ag/agl/aglw/aquastat/irrigationmap/index.stm).
This map presents areas that are ?equipped for irrigation? but not necessarily
irrigated (Siebert et al. 2005; Siebert et al. 2002; Siebert and Döll 2001; Döll and Siebert 1999, 2000).
The map is produced using irrigated area statistics from various nations. GIS and national statistics
based irrigated area maps are also available for individual nations such as India?s
CBIP maps which may have following limitations.?. First,
extrapolating the statistical numbers to spatial domain can be a rough
approximation of the actual location of the irrigated areas. As a result we may
have an entire state such as Washington state in USA having <5 percent
irrigation with no indication on which specific areas this irrigation takes
place. Second, irrigated area statistics provided by different countries have
various inconsistencies. There is a tendency to believe in ?official?
statistics as right one. However, a cursory look at these data often highlights
numerous inconsistencies. For example, the irrigated areas of the 29 Indian
states had 99 percent correlation between areas of 1995-96 and 2000-01. This
simply implies that the same numbers from previous years have been copied in
subsequent years. Third, it does not account for the intensity (gross area) of
irrigation. Irrigated area maps and statistics from various Nations have their
own limitations. For example, the Central Board of Irrigation and Power (CBIP,
1995) of India calculates irrigated areas based on the irrigated command area.
Our studies at 500-m resolution, currently in progress and within the scope of
GIAM project, showed very significant proportion of the command area are left
fallow at any given period of time. Further, within the command area
boundaries, there are other classes: ground water irrigation, rainfed
croplands, and other land use\land cove. The command area maps help establish
?equipped area? but not actual area. The gap between ?actual? versus ?equipped?
can be significant. Another source of inconsistency concerns the cropping
intensity which varies from year to year and among systems and regions.
The FAO/University of Frankfurt (FAO/UF) study estimates area equipped for irrigation to be 274 Mha or about 16
percent of the total croplands (1.5 billion ha) around year 1995. The pixel resolution presented by FAO/UF is based on sub-national statistics and variable scale maps and administrative units (Siebert et al., 2005).
Irrigated area is also estimated,
rather coarsely, in global land use classifications derived from remote
sensing, which have usually focused on other objectives, such as forestry,
rangelands and rainfed croplands. Examples include USGS 1993 (Loveland, et al.,
2000), GLC 2000 (Bartholome“ and Belward,
Global Forest Cover (DeFries et al., 2000a and 2000b).
Settled agriculture began about
10,000 years ago. There are many examples of irrigation dating back to at least
4000 B.C. in great ancient civilizations in the Nile, Euphrates, Indus, and Ganges (Postel, 1999). Irrigation was practiced extensively in the ancient world
in the Tigris and Euphrates by Sumerians, Babylonians and Mesopotamians about
2000 to 6000 years ago, and by the Harappa and Mohenjedaro civilizations in the
Indus valley about 4000 years ago. In the Nile Delta, there has been a near
continuous practice irrigation over 6000 years and large scale systems have
been continually expanded in China for up to 4000 years, for example in
Dujiyangyan, in Szechuan, which now covers an near contiguous area of nearly 1
Historical estimates of global
irrigated area begin with 8 million hectares in year 1800, rising to 95 Mha in
1940, to the current ones. About 60 percent irrigation is found in six
countries: India (21.7 % of the total World?s irrigated area); China (19.4%); USA (7.9%); Pakistan (6.6%); Iran (2.8 %); and Mexico (2.4%) (Droogers,
2002). These countries also have the highest proportions of irrigation relative
to total cultivated area, for example: 50.1% for India, 49.8% for China, 21.4% for USA, 17.2% for Pakistan, and 7.3% for Iran (Postel, 1999).
Satellite sensors potentially offer
a consistent, continuously updated, timely and increasingly free resource that
meets high scientific standards, such as MODIS and SPOT Vegetation which
respectively have 250 meter to 1-kilometer spatial resolutions with global
coverage every day. These data are backed by numerous high quality secondary spatial
data such as SRTM digital elevation models, Landsat, SPOT and ASTER high
resolution data and global time-series of precipitation and other climatic
The International Water Management
Institute (IWMI) initiated a Global Irrigated Area Mapping (GIAM) project in
year 2002 (see Droogers, 2002, and Turral, 2002) supported by the Comprehensive
Assessment (CA) on Water Management in Agriculture and by IWMI.
The main motivation to develop
the IWMI map lies in the potential of a wide range of increasingly
sophisticated remote sensed images and techniques to reveal vegetation dynamics
- define more precisely the actual area and spatial distribution of
irrigation in the World;
- elaborate the extent of multiple cropping over a year,
particularly in Asia, where two or three crops may be planted in one year,
but cropping intensities are not accurately known or recorded in secondary
- develop methods and techniques for a consistent and un-biased
estimates of irrigation over space and time for the entire World.
4. Data used in creation of the IWMI?s Global irrigated
Time series data potentially
allows the often distinct dynamics of irrigated agriculture to stand out from
other land use, but there are many confusing situations: for instance in the
tropics, where rice may be mainly rainfed in the monsoon season but receive
some irrigation, and is followed by one or more dry season crops which may be
completely irrigated. In tropical environments, there is generally a high
degree of land cover the whole year round and everything is ?green?, making
precise definition of irrigated crops more difficult, especially if relatively
coarse scale imagery is used.
In this analysis, we make use of
as much freely available data as possible. AVHRR and MODIS data are relatively
coarse scale, with resolutions from 10-km down to 250-m. Compiling a MODIS data
set for the world at 500-m or 1-km over time (e.g., 8-day or monthly for
several years) requires enormous computer storage and extremely high end
processors that are expensive. The longest multitemporal series of remote
sensing data with global coverage is AVHRR 8-km (re-projected to 10-km). However,
since this resolution is coarse, we have combined a three year monthly time
series of AVHRR 10-km from 1997-1999 with a 1-km SPOT Végétation mosaic of the
world for 1999. A summary of the data used, and its main processing chain is
summarized in Figure 1.
The process starts with a number
of publicly available data sets, which are processed into one large 159-layer
time series file, known as a mega-file. The time series analysis is conducted
on the mega-file and is described in sections 4 and 5. DEM, temperature and
rainfall data is combined into the megafile to allow segmentation of a set of
masks (Figure 1) of different characteristic regions of the world which are
analysed separately and then combined in the class naming and area calculation
steps. A number of other data sets (Figure 1) are used to provide contextual
and detailed information to assist in identifying, separating and aggregating
All input data, mega-file and outputs are stored in the IWMI
Data Storehouse Pathway (IWMIDSP), an on-line archive which stores all remote
sensing and GIS data collected by IWMI. The site can be accessed at: http://www.iwmidsp.org
The mega-file used for the IWMI
Global irrigated area map (GIAM) consisted of 159 data layers (Figure 1). This
consisted of: 144 AVHRR 10-km layers from 3 years (12 layers from 1 band per
year * 4 bands including an NDVI band * 3 years), 12 SPOT vegetation 1-km layers
from 1 year, and single layers of digital elevation model (DEM) 1-km, mean
rainfall for 40-years at 50-km, and AVHRR derived forest cover at 1-km. The 159
band mega-file data layers were all retained at a common resolution of 1-km by
re-sampling the coarser resolution to 1-km.
Figure 1. Processing chain for the Global
irrigated area map (GIAM).
Figures 2 and 3 illustrate
various types of data present in mega-file. The drop-down menu of bands shows
how the layers are ordered. The following sections provide a brief description
of each of the data sets, which are summarized in detail in Tables 1 and 2.
4.1 Primary remote sensing data sets and masks
4.1.1 AVHRR data characteristics
time-composite NOAA AVHRR 0.1 degree data are obtained from the NASA Goddard
The ?Pathfinder? data set has gone through many stages of calibration and
re-calibration (Smith et al. 1997; Rao, 1993a; Rao, 1993b and Kidwell, 1991)
making it a high quality science dataset. The original scaled 16-bit and 8-bit
data have been converted to three primary variables: (a) at-ground reflectance,
(b) top of atmosphere brightness temperature, and (c) NDVI. These parameters
were derived using calibration parameters in the following four equations:
|Reflectance (percent) = (Band 1 scaled DN in 16-bit radiance – 10) *
|Reflectance (percent) = (Band 1 scaled DN in 16-bit radiance ? 10) * 0.002
|Band 4 brightness temperature (° Kelvin) = (Band 4 scaled DN in 16-bit + 31990)*0.005
|Normalized difference vegetation index (NDVI) = (SNDVI ? 128) * 0.008
Bands 1 and 2
have been processed through standard radiative transfer equations and been
corrected for Rayleigh atmospheric scattering (Gordon et al. 1988). Moisture
absorption effects have been corrected using data from the Total Ozone Mapping
Spectrometer (Fleig et al., 1983). The resulting reflectances were then normalized
for solar illumination (NGDC, 1993). The band values and NDVI distortions due
to external forcing (e.g., stratospheric aerosols and satellite orbit
degradation) are serious concern and need to be addressed (Kogan, 2001). For the
thermal channel, first the atmosphere radiance was calculated and converted to
brightness temperature using a Planck function equivalent lookup table, based
on the response curve of each channel.
A critical issue in long-time
series is data normalization. Many factors lead to variations or shifts in data
calibration that include, but are not limited to: sensor degradation; changes
in sensor design; satellite orbital characteristics; atmospheric effects;
topographic effects; and sun elevation. AVHRR Pathfinder data has gone through
many processing steps and most of these effects were already corrected prior to
use in this analysis.
The monthly maximum value composite
(MVC) data from 1981 to 1999 is stored in a single mega-file of 239 bands, and
this was also used to generate animations of NDVI and skin temperature to
assist in understanding vegetation dynamics and identify irrigated area. A
subset of three years of this data (1997-1999) was incorporated into the
irrigation mapping mega-file.
Table 1. Characteristics of the Satellite sensor and secondary datasets
used in mapping Global irrigated areas. These datasets
were Compiled into a 159-band layer stack1,2.
number3 Or primary source
of bands and radiometry
(#; one per month)1
final format Z-scale
(percent: for reflectance)
|Satellite sensor data
|Band 1 (B1)
||0.58 - 0.68
|Band 2 (B2)
|Band 4 (B4)
| unitless, 8-bit
||DCW, DTM, and
to + 1
||Mean of monthly
|class names, 8-bit
2. Other data used in conjunction
with the megafile
1, 2, NDVI
||same as above
|2. SPOT 1-km2
|3. JERS SAR 100-m
animations of the irrigated area classes were run for the entire AVHRR
time-series data to help understand the change history of the class. There was
data for 239 months in 19 years (July 1981- September 2001). September-December
1994 data was not acquired due to failure of the satellite.
all data were calibrated and normalized by data provider (see section 4.1.1).
Figure 2. Mega-file used in GIAM. The mega-file of 159
layers of data and consist of 144 AVHRR 10-km monthly layers from 3 years,
SPOT monthly layers from1999 year, single layer of DEM, mean annual rainfall
from 40-years, and forest cover.
3. Primary and secondary data sets used in the Mega-file.
4.1.2 SPOT data characteristics
The SPOT Végétation (SPOT VGT) 1-km
data has 4 wavebands: blue (0.43-0.47 µm); green (0.61-0.68 µm); near-infrared
(NIR) (0.78-0.89 µm); and short wave infrared (SWIR) (1.58-1.75 µm). There is
10-day synthesis of SPOT VGT data that can be downloaded free of cost for the
entire world (http://free.vgt.vito.be/).
A single year monthly SPOT VGT NDVI data for 1999 was used in this study.
Similar corrections have been
made to the SPOT VGT data by the SPOT Image production team, including scattering
and moisture absorption using radiative transfer models. Cloud and snow were
detected using a multivariate thresholding technique and neural networks using
data from all 4 VGT wavebands (Lissens et al., 2000). Cloud shadows were
detected using geometrical model as described by Lissens et al. (2000). The
10-day synthesis is performed using daily data using maximum value compositing
4.1.3 Mask data
Secondary data sets in the mega-file
are used to segment the world into characteristic regions based on rainfall,
elevation, temperature and known forest cover. For example, in areas where
temperatures are less than 280K, it is unlikely that there is any vegetation
and little chance of there being any irrigation.
30 1-km DEM
The GTOPO30 is derived from eight
sources consisting of digital terrain elevation data or DTED (50% of global
coverage), digital chart of the World or DCW (29.9%), USGS 1-degree digital
elevation models (6.7%), army map service maps (ASM maps) at 1:1,000,000 scale
(1.1%), international map of the world (IMW maps) at 1:1,100,000 scale (4.7%),
Peru map at 1:1,000,000 scale (0.1%), New Zealand DEM (0.2%), and Antarctic
digital database (8.3%) (Tucker et al., 2004; USGS, 1993; Verdin and Greenlee,
1996; and Verdin and Jenson, 1996). The vertical accuracy of the component DEM
data varies significantly from source to source. Accuracies were 30-m for DTED,
160-m for DCW, 30-m for USGS DEM, 250-m for ASM maps, 50-m for IMW maps, 500-m
for Peru map, 15-m for New Zealand map, and highly variable for Antarctic
digital database. In this paper, GTOPO30 data were used to segment the image
based on elevation gradients.
4.1.5 CRU precipitation
and temperature data
The 40-year (1961-2000) monthly, 0.5 degree, interpolated
rainfall and temperature data were obtained from Dr. Tim Mitchell of the Climate
Research Unit (CRU), University of East Anglia, UK (Mitchell, et al., 2003) (http://www.cru.uea.ac.uk/~timm/index.html).The data have been
converted to ESRI GRID format at IWMI and mean monthly precipitation and
temperature for 40-years were computed for each pixel and added to the mega-file.
4.1.6Forest cover data
Forest cover was derived from the
1992 AVHRR 1-km data by university of Maryland that used a continuous fields
approach (rather than discrete number of classes) using a linear mixture model
approach (see DeFries et al. 2000a and DeFries et al. 2000b). This dataset was
used to mask areas of very high forest cover, which implies the land is not
available for cultivation or irrigation.
4.2 Secondary data sets
4.2.1 JERS-1 SAR derived forest cover
Rainforests may contain
fragmented irrigation along with shifting cultivation and clearance for
livestock production. Mapping irrigated areas in rainforests is more complex
than in other parts of the world as a result of forest fragmentation and
significant cloud cover. Even the monthly AVHRR and SPOT VGT MVCs contain some
cloud cover over rainforests and irrigated fragments are difficult to discern
at 1 to 10-km. We obtained 100-m resolution JERS-1 SAR tiles (
http://southport.jpl.nasa.gov/GRFM/) for South America and Africa to assist in mapping major rainforest areas at higher resolution.
Unfortunately, well processed JERS SAR images are not readily available for Asia and hence could not be used.
4. JERS-1 SAR 100-m image tile mosaicks for the Central Africa.
The rainforests of the Africa and the Central America were studied using JERS-1 SAR
for two periods in 1995-1996.
The Japanese Earth Resources
Satellite-1 (JERS-1) Synthetic Aperture Radar is a L-band (24.5 cm wavelength)
imaging radar with initial full-resolution of 18-m, that is processed to 100-m,
mosaicked and made available for the entire contiguous rainforests of Amazonia and Central Africa. The JERS SAR antenna has a median look angle of 35 degrees.
The Amazon basin was imaged by
JERS-1 during a low flood period from September-December 1995 and in a high
flood time from May-August 1996. The Central and West Africa rainforests images were obtained for January-March 1996 and October-November 1996. Over 20 million
square kilometers of the rainforests are covered by these images (Saatchi et
al., 2000, Saatchi et al., 2001). The JERS SAR image tiles were mosaicked into
single files for Central Africa (Figure 4) and Amazon using ERMapper 6.5, at IWMI.
The 8-bit JERS SAR data of rainforests were analyzed separately and fused with the overall classification results from other
Landsat 150-m GeoCover
re-sampled the 8,500 ortho-rectified Landsat ETM+ ?GeoCover? tiles that had
been produced by the EarthSat Corporation (http://www.earthsat.com),
funded by NASA (Tucker et al., 2004). The original images are free from the
USGS EROS data center and the University of Maryland (http://glcf.umiacs.umd.edu/index.shtml).
The re-sampled images have a pixel resolution of 150 m compared with the
original pan-sharpened size of 15m. GeoCover is the most positionally accurate
image set covering the entire globe and shows maximum greenness. They offer a detailed
?zoom-in? view of any part of the world (Figure 5) and are used to provide
contextual information and pseudo ?ground-truth? by geo-linking to the class
maps to identify and label classes.
5. Landsat ETM+ 150-m images of the World as ?ground-truth?.
ETM+ (GeoCover 2000) orthorectified images for the nominal year 2000 at 150-m resolution were used as a ?ground-truth?.
4.2.3 Google Earth Dataset
Google Earth (http://earth.google.com/) contains an
increasingly comprehensive image coverage of the globe at very high resolution
0.61-4m, allowing the user to zoom into specific areas in great detail, from a
base of 30m resolution data, based on GeoCover 2000. This assists:
- Identification and labeling the
- Area calculations (section 7); and
- Accuracy assessment of the classes
(see section 8).
For every identified class, 20-50 sample locations were cross-checked
using Google Earth.. The indicative GIAM class name can be updated according to
the dominant class identified at high resolution within the sample area. The
process also helps to identify mixed classes. Google Earth data is used as a
substitute for groundtruth, although images may in fact be snapshots of
cropping systems taken at different times. The very high resolution data
has some advantage over real groundtruth in that it provides information on a much larger areas, and therefore more representative area than is normally
sampled directly on the ground.
Precise knowledge of the real
situation on the ground is essential to the interpretation of all remote
sensing products for: training; class identification and naming; and for
accuracy assessment. Clearly, it is hard to obtain sufficient detailed
groundtruth data to cover the whole world, but there is a growing amount of
publicly available data thanks to the internet. IWMI actively collects
groundtruth data for more localized projects, and undertook two major
groundtruth campaigns in India in 2003. Four more were completed in 2005, one
in Southern Africa, one in Central Asia and two in India. The project team is
working to continually expand the geographical scope, range and detail of
groundtruth data available and all ground truth data is archived in the IWMI?s
data storehouse pathway or IWMIDSP (http://www.iwmidsp.org).
There are two global archives of GT data, one collected by IWMI and its staff and one using public domain data from the degree confluence project (http://www.confluence.org/).
4.2.5 Groundtruth at IWMI: data collected in field campaigns
ground-truth (GT) data were collected by IWMI specifically for irrigated area
mapping (see for example, http://www.iwmidsp.org
and also in Thenkabail et al. 2005a, Thenkabail et al. 2005b, Biggs et al. 2005).
The India GT mission covered about 12,000 kilometers and collected data from
393 specific points in September-October 2005, which is the peak of the monsoon
crop season (Kharif, July-October). The Ganges basin field campaign was
conducted from October 1 to 22, 2003 to coincide with the peak crop growth
stages in Kharif. The Krishna River Basin data were collected from October 13 to 26, 2003. Data was gathered for 144 sample sites in Krishna and 196 sites in
the Ganges region, covering all available land types. The precise locations of
the samples were recorded by GPS in the Universal Transverse Mercator (UTM) and
the latitude/longitude coordinate system with a common datum of WGS84 (see
location the following data were recorded (Thenkabail et al. 2005a):
Figure 6. Ground-truth (GT) data collected
- LULC classes:
levels I, II and III of the Anderson approach.
- Land cover
types (percentage): trees, shrubs, grasses, built-up area, water, fallow
lands, weeds, different crops, sand, snow, rock, and fallow farms.
- Crop types,
cropping pattern and cropping calendar for Kharif, Rabi (winter or dry
season cropping period from November to March) and interim seasons.
source: rain-fed, full or supplemental irrigation; surface or groundwater.
photos hot linked to each ground truth location.
Groundtruth data assembled from multiple locations and times by IWMI projects and staff.
Data accumulated by IWMI and its staff
amount of groundtruth data has been collected on research projects conducted by
the primary author over the last decade and was added to the archive of other
data already held by IMWI. The data is sourced from India, Sri Lanka, Syria, West and Central Africa, South Africa, and Central Asia (see Figure 6).
4.2.6 Public domain groundtruth ? the Degree Confluence
The Degree Confluence Project
(DCP) (http://www.confluence.org/) is
an organized sampling of the entire World at every 1 degree latitude and
longitude intersection. It is a voluntary effort and close to 4000 confluence
locations have already been contributed. The confluence points include precise
latitude, longitude and a digital photo of land cover. These were converted to
proprietary GIS formats and added to the DSP in a separate archive to preserve
We used DCP data to interpret land use at each
location, based on the digital photo and added the complete set to a GIS (see
Figure 7). These photos were used as part of the accuracy assessment and for
7. Ground-truth data of the World from the Degree Confluence Project (DCP).
4.3 Other data sets for comparison purposes
A number of existing global land
use/land cover (LULC) products were used in preliminary class identification
and labeling process. These included USGS LULC (Loveland, et al., 2000), USGS
seasonal LULC (Loveland, et al., 2000), GLC2000 (Bartholome“ and Belward, 2005),
IGBP (IGBP, 1990), and Olson eco-regions of the world (Olson 1994a, 1994b).
These data supplemented/complemented the ground-truth data during the
preliminary class identification and labeling processes. The characteristics of
these LULC classes are briefly mentioned here and for further detail the reader
is referred to peer-reviewed publications.
The Global Land
Cover 2000 (GLC2000) data set was derived using data from SPOT 1-km resolution
Végétation Instrument (Bartholome“ and Belward, 2005, Agrawal et al., 2003).
The 10 day synthesis data from November 1, 1999 through December 31, 2000 were used for the classification (http://www.gvm.sai.jrc.it/glc2000/Products/).
The Global Land Cover characteristics database was
developed on a continent-by-continent basis using 1-km, 10-day AVHRR data
spanning April 1992 through March 1993 (Loveland et al., 2000). The same
primary data was used in the Global USGS LULC, seasonal USGS LULC, and IGBP
provide global 94 unique ecosystem classes for the Globe (Olson 1994a, 1994b) (http://edcdaac.usgs.gov/glcc/globe_int.html).
This approach was developed in the mid-1980s and did not use any remote sensing
information. For easy convenience, all these land cover products are made
available in standard image processing formats (e.g., ERDAS Imagine) in IWMIDSP (http://www.iwmidsp.org).
An overall summary of the methods
and analytical techniques is shown in Figure 8. The basic process involves
segmenting the world into characteristic regions that are easier to analyze and
then performing an unsupervised classification on each segment, containing all
the 159-band information from the AVHRR time-series and the single year of SPOT
VGT data. Identification of the resulting classes is performed using a suite of
new techniques to interpret vegetation dynamics in multitemporal series, which
are explained in more detail below. A number of classes could not be clearly
identified, and so were subdivided and classified using simple decision trees
and ?ground-truth? data sourced from GeoCover 150-m and other secondary information (Tucker, et al. 2005). This resulted in generic class map of 951 ?unique? classes. As
far as possible, class naming was harmonized with earlier Global Land Cover
classifications. Irrigation classes were then derived by aggregation of similar
irrigated land use in the generic map, resulting in a 28 irrigation class map (GIAM10km-28
Classes). This map is used to estimate irrigated crop areas in each of three reference
seasons (see section 8). A further aggregation of this map into 8 broad
irrigated area classes of the World (GIAM10km-8 Classes) gives a more visually
friendly presentation, with class names that are more familiar to irrigation professionals.
5.1 Image segmentation
5.1.1 Mega-file of segments
The original 159 band mega-file
was converted into a megafile of segments, each with its own set of 159 bands
(see Figure 1). The seven global masks created are listed below and then brief
examples of each one are presented in Figure 9 and 10. The Global masks are:
- Precipitation less
than 360 mm per year (PLT360);
greater than 2400 mm per year (PGT2400);
- Temperature less
than 280 degree Kelvin per year (TLT280);
- Forest cover
greater than 75 percent canopy cover (FGT75);
- Special forest SAR
- Elevation greater
than 1500 meter (EGT1500); and
- All other areas of
the World (AOAW).
The segment with less than 360 mm
per year identifies areas where any green vegetation has a very high likelihood
of being irrigated, since average evaporation rates of 30mm per month, however
distributed in reality, will be considerably less than evaporative demand. This
segment will mainly identify arid, semi arid areas and deserts, as shown in
Figure 9. By contrast, the
segment with rainfall over 2400mm per year mainly identifies the rainforest areas of the world, although there are considerable areas of irrigation within the SE
Asian lands. Where temperature is less than 280K on average, it is too cold for
agriculture, and irrigation is not likely to be found there. However, some
northern hemisphere areas have low average temperature but short summer seasons
in which supplemental irrigation is actually practiced.
Figure 8. Summary of analysis to determine irrigation and
land use classes (Part 1).
Figure 8. Summary of analysis to determine irrigation and
land use classes (Part 2).
Figure 9. Precipitation less than 360 mm segment (PLT360-segment).
These arid or
semi-arid areas provide distinct contrast between areas with and without
Figure 10. Forest density greater than 75 percent (FGT75-segment).
These areas have low probability
except in rare fragments of slash and burn.
If forest density is greater than
75%, it is also rare that there will be any irrigation, due to high rainfall
and limited infrastructure. There is likely to be slash and burn agriculture in
small fragments. This mask in complemented by a special rainforest mask derived from the JERS-1 SAR imagery, in order to better identify other land use
fragments at higher resolution within the rainforest areas, including where
there might be irrigation.
There is a lower likelihood of
irrigation above 1500m elevation, although there are certainly hill irrigation
systems in the Andes, Himalayas and in the Philippines at higher elevations. Forest is a likely land cover, but should be separable from irrigation and agriculture due
to its continuous vegetation signature. Finally, the segment ?all other areas
of the world? focuses on where there are few bio-physical constraints to
irrigation and shows where we are most likely to find it in various forms.
segment is processed using unsupervised ISOCLASS K-means classification (Tou
and Gonzalez, 1974, ERDAS, 2005). This calculates class means evenly
distributed in the data space and then iteratively clusters the remaining
pixels using minimum distance techniques. Each iteration recalculates the means
and reclassifies pixels with respect to the new means. Iterative class
splitting, merging and deleting is done based on input threshold parameters
(see ERDAS, 2005). All pixels are classified to the nearest class unless a
standard deviation or distance threshold is specified, in which case, some
pixels may be unclassified if they do not meet the selected criteria. This
process continues until the number of pixels in each class changes by less than
the selected threshold or the maximum number of iterations is reached.
5.2.1 Preliminary Class identification and naming
On completion of an unsupervised
classification, it is necessary to identify what the classes are and label them
accordingly. In more localized applications, it is common to undertake ground-truth
after a preliminary unsupervised classification, which identifies
characteristic land units for investigation and this was done for the IWMI field campaigns in India. However, at global scale this is not possible, and a combination
of techniques is employed to first group classes based on the similarity of
their time series behaviour, then identify in more detail what they are through
understanding the spatial-temporal variations in reflectance and cross
referencing to higher resolution images (GeoCover 150; Tucker et al., 2005),
existing GIS, maps and groundtruth data.
5.2.2 Spectral Matching Techniques
Time series of NDVI or other
metrics are analogous to spectra, where time is substituted for wavelength. Considerable
research effort has been made into hyperspectral imagery analysis and this
yields a number of promising avenues, developed here, for the analysis of time
series. Spectral Matching Techniques (SMTs) have mostly been applied to
hyperspectral data analysis of minerals (Homayouni S. and Roux M., 2003; Shippert, P. 2001, Bing et al. 1998; Farrand and Harsanyi,
1997; Granahan and Sweet, 2001; and Thenkabail et al. 2004c and 2004d).
The principle in spectral
matching is to match the shape, or the magnitude or (preferably) both to an
ideal or target spectrum (commonly know as a pure class or ?end-member?). The
time series signatures of irrigated crops across the globe can match (tropics)
or be out of phase (tropics and southern hemisphere). In this work we used a
combination of qualitative and quantitative approaches to identify target
spectra and match others to them. The selection of the target spectra is
governed by groundtruth and GeoCover2000 to identify consistent LULC and
irrigated areas classes. Example locations, where we know the signatures
precisely, for ideal target spectra used in this work are shown in Table 3.
The locations of the ideal target spectra for 7 irrigated area classes.
Type of Irrigation
1. Delta irrigation:
Krishna Delta (India)
15 59 17.81 N
80 57 07.55 E
22 10 22.31 N
90 44 14.44 E
Yellow River Delta (China)
37 00 15.93 N
118 12 16.56 E
Mekong River Delta (Vietnam)
10 30 29.09 N
105 12 02.16 E
Nile Delta (Egypt)
31 08 05.65 N
31 03 30.10 E
2. Large Scale
1. Tungabhadra Reservoir (Krishna River Basin)
15 41 31.43 N
76 41 50.54 E
2. Nagarjuna Sagar (Krishna River Basin)
16 29 29.87 N
80 52 00.95 E
Ganges River Basin (Haryana)
30 43 39.20 N
76 27 41.24 E
Rechna Doab (Indus River Basin)
32 02 10.58 N
74 33 28.86 E
Indus RB (Lower part)
25 39 40.32 N
68 34 57.51 E
39 27 35.96 N
121 48 31.89 W
South East Australia
29 39 27.00 S
116 01 57.34 E
4. Canal irrigation;
1. Ganges River Basin (Upper Uttar Pradesh)
27 06 22.85 N
78 40 08.24 E
5. Centre Pivot
37 42 54.93 N
106 03 45.28 W
47 09 01.26 N
119 44 31.51 W
42 39 48.81 N
59 00 47.76 E
6. Flood irrigation
(River flood plains);
Ganges River Basin (Patna)
25 32 32.88 N
85 08 25.13 E
Brahmaputra flood plains (Assam)
26 07 44.84 N
90 55 47.45 E
43 09 06.11 N
97 49 40.99 W
35 53 06.71 N
37 03 28.33 E
attempted to use Modified Spectral Angle Similarity (MSAS) (Shippert, P. 2001, Homayouni S. and Roux M., 2003,
Farrand and Harsanyi, 1997, Schwarz and Staenz, 2001, Thenkabail et al. 2005b)
which measures hyperspectral angle between spectra of any 2 classes or between target
and sample class spectra. However, the practical implementation of this was
troublesome (see also Thenkabail et al., 2006), often providing uncertain
results, and so is not discussed further.
Qualitative spectral matching
spectral matching is often performed before quantitative approaches (e.g.,
Figure 11. It provides a preliminary indication of which classes group together
and which stand apart. Indeed the classes that match up through: (a) shape
only, and/or (b) magnitude only, and/or (c) both shape and magnitude, are
identified visually. When two classes, such as continuous irrigation and
forests, match and provide high quantitative correlations, it is essential to
plot both classes with reference to their spatial location using ground truth or
Time-series AVHRR 10-km profile of spectral classes is illustrated for
AOAW-segment. The AOAW-segment initially had 350 classes. The plot of some
of these classes highlights the spectral characteristics of each class. A
quantitative approach to determine which of these classes match is performed
through SCS R2-squared (e.g., Table 4).
Quantitative spectral matching
In this study, Spectral Correlation Similarity (SCS) has
been applied to match the shape of any class to the selected target class.
Spectral Similarity Value (SSV) has been used to determine the match of both
shape and magnitude. SCS is defined as follows and is based on Pearson?s correlation
to an NDVI time series (SAS, 2004, van der Meer and Bakker, 1997):
= NDVI time-series (i=1 to n) of the target class;
= Mean of the NDVI time series of the target class;
= NDVI time-series (i=1 to n) of any other class;
= mean of the NDVI time-series (i=1 to n) of any other class;
standard deviation of target class NDVI time-series (i=1 to n); and
standard deviation of NDVI time-series of other class.
The range of
Pearson?s Correlation Coefficient ()
normally lies between -1 and +1, but negative values have no meaning in this
application. The higher the value, the greater is the similarity.
defined as follows (Homayouni S. and Roux M., 2003, Granahan and Sweet, 2001):
Euclidian (shortest) distance between two points
Pearson?s Correlation Coefficient as defined above.
range of SSV is from 0 to 1.415 and the smaller the value, the greater is the
The SCS R2-value matrix of spectral classes. The spectral
correlation similarity (SCS) R2 value matrix is illustrated for 2-spectral
classes. Those highlighted in cyan are highly correlated (R2-value
greater than 0.92) with 11 other spectral classes. Similarly spectral classes highlighted
in yellow and orange are moderately correlated with thresholds of 0.9-0.92 and
0.84-0.90 respectively. SCS is the first step towards grouping classes,
providing a strong indication of which classes have similar characteristics.
of spectral matching is illustrated beginning with a plot of multiple time
series and two selected target series in Figure 11, which are characteristic of
two irrigated crops per year in the Indian subcontinent. Figure 12 shows the
results of grouping similar spectra for double crop irrigation, continuous
forest cover, and bare or fallow soils. The process involves matching spectra
of classes, grouping classes with similar spectra, and then identifying and
labeling classes using Figure 8 protocol. The results of a cross correlation of
64 time series is shown as a subset in Table 4 and shows, for example, that
Class 1 is highly correlated with five other classes (2,19, 20, 42 and 44).
Similar matrices were calculated for each segment classification. Figure 13 shows
where the spatial similarity of classes across the globe is determined with
reference to known ground condition and high values
in the matrix.
Identifying similar irrigated classes using spectral matching. Spectral
matching in combination with ground truthing and ideal spectra helped group similar
irrigated (shown in dark green, for classes 25, 26, and 27). The same logic was
used to group: forests (sown in light green; class numbers 1, 2, 3, 4 and 5),
Savanna/Croplands mix (Orange; class 50, 59, 60, 67, 74), and Barren/Deserts (shown
in blue; classes 10 to 15).
extraction and geographical location of similar classes is shown in a more
pictorial way in Figure 13. Figure 14 illustrates where there is good shape
similarity but poor correlation in magnitude, which indicates that the classes
should be separated. Such time series were compared with target spectra from
ideal locations (Table 4) and those with low values of SSV were grouped
The use of
SSV is illustrated taking three distinct types of irrigated areas-major
irrigation, supplemental irrigation, and delta irrigation (Figure 16). The
target time series is selected for a ?pure? location for the class depicting
near ideal conditions. The unsupervised class spectra are sourced from a much
larger number of pixels, and hence depict an average situation. The target
classes in this example are: (a) major irrigation in Ganges basin (target in
red, unsupervised grouped class series in magenta), (b) supplemental irrigation
from mid-west USA (centre pivot sprinkler) and Syria (groundwater) (targets in
light blue and actual unsupervised class series in dark blue), and (c) delta
irrigation from Bangladesh (ideal series in light green, unsupervised class
series in dark green).
The process of combining classes in spectral matching techniques (SMTs) is
illustrated. First, the SCS R2-values are determined for
a matrix of classes. The time-series spectra of classes with high SCS R2-values
are then matched. Grouped classes are investigated further using all other
types of information including groundtruth. This leads to distinct groups such
as: boreal forests and tropical forests. Finally, the classes of similar types
are color coded.
The process of spectral matching techniques (SMTs) is illustrated. The 17 classes
considered in Figure 12 are further refined by quantitative and qualitative
SMTs that lead to 3 distinct groups.
small SSV of 0.22 and a moderately high SCS (or
SCS R2) value of 0.84 between the major irrigation target series and
unsupervised classes spectra show that both the shape and magnitudes match. For
supplemental irrigation the match is even better: a very small SSV of 0.16 and
a very high SCS R2 value of 0.97 When dissimilar classes like major
irrigation and supplemental irrigation are matched, there is an obvious
mismatch, captured by high SSV and low SCS R2 values.
Figure 15. The spectral similarity value (SSV) to match spectra. In this figure,
unsupervised class spectra are compared with ideal spectra of distinct
irrigated classes: (a) major irrigation in Ganges basin (ideal spectra in red,
unsupervised grouped class spectra in magenta), (b) supplemental irrigation
from mid-west USA (pivot sprinkler) and Syria (underground water) (ideal
spectra in light blue and actual unsupervised class spectra in deep blue), and
(c) delta irrigation from Bangladesh (ideal spectra in light green,
unsupervised class spectra in deep green). The smaller the SSV, the greater the
match in shape and magnitude.
Earth as a resource for class naming
Once the classes
are grouped by spectral similarity, each one is investigated for their class
characteristic by taking 20-50 sample points on Google Earth spread across the world
(Figure 16). If there is a overwhelming evidence that the class falls into a
particular category, an indicative name is assigned. The interpretation of a
class is based on visual indicators such as shape (e.g., central pivot circles),
size (e.g., reservoir size for large and small scale), pattern (e.g., contiguous
farms), and texture (e.g., smooth texture of a farm compared to rough texture
of a forest). The process is repeated for every class in a group. If the Google
Earth sample points for a class indicate a mixed land use/land cover, then the
class is further processed either through decision trees or is re-classified,
or GIS spatial modeling is applied to derive homogeneous classes.
Google Earth ?zoom in? views to identify a class. One preliminary class is spread
out across the world. The class was investigated using 50 Google sample points
that were randomly chosen. The figure shows the spread of the class across the
world and Google Earth hi-res image at 2 locations: center pivot ground water
irrigation in the USA and surface irrigation in Sudan.
points (e.g., yellow points also called ?place marks? in figure 16) were used
in identifying and providing indicative class labels in generic 951 class GIAM10km
5.4 Advanced techniques for class identification
A summary of the application of
brightness, greenness and wetness characteristics applied to multi-temporal
imagery is discussed recently by Thenkabail et al (2005) with respect to
irrigation mapping in the Ganges Basin in India. A brief review is given here.
2-dimensional near-infrared vs. red band spectral reflectivity plot of
unsupervised classes is referred to as a brightness-greenness-wetness (BGW)
plot (Figure 17). The BGW plots help determine whether a class is: (a) green,
(b) bright, (c) wet, or (d) somewhere in-between these classes. Classes that
occupy green area have high NIR reflectivity and low red reflectivity.
Typically, these areas are forests, agricultural lands, and natural vegetation.
Classes that occupy bright areas have high NIR and high red reflectivity. The
land use/land cover (LULC) categories of these classes are likely to be
open/barren areas, sparse vegetation, dry vegetation, clouds and built-up areas.
Classes that occupy wet areas have low NIR and low red reflectivity. These classes
are likely to be wetlands, moist lands, water bodies, cloud shadows and swamp
forests. The classes that are in between have different combinations of these
broad LULC classes.
Figure 17. Brightness-greenness-wetness (BGW) plot fundamental principles.
plots provide clear and useful information on class dynamics over time and are a
very helpful tool in identifying and labeling a class.
5.4.1 Brightness, greenness and wetness for a single
Single date BGW plots do not
capture the dynamics of a time series, but they show why NDVI (as the most
commonly used metric of vegetation) can be low but represent different land
uses ? for instance, when it is either wet or dry, or closer or father away
from the soil line for the date in question. Similar plots are then established
for individual dates of the time-series data. The
class location in 2-D space is dependant on time, and classes such as
continuous cropping, forests and deserts do not move much. Bigger changes over time
can be seen in classes such as rainfed crops, irrigated crops and rain
dependant grasslands. More subtle changes occur when, for example, a crop
develops from an early to a maximum vegetative growth stage, in a short time period. This is more clearly illustrated
in multi-date plots.
dynamics of brightness, greenness and wetness
plots could be created for every image date in a time-series. However, plotting
for a series every few months, to highlight different cropping seasons provides
sufficient dynamic information to characterize different land uses. For
example, the band 1 vs. band 2 spectral reflectivity of AVHRR data for
unsupervised classes are plotted over the peak months of the cropping seasons: January
(blue), May (green), and September (red) in a given year on a single plot. This
allows tracking a particular class through the peak months of each of 3
seasons. We can see that class 10 has very considerable changes over the
18, the spectral reflectivity of AVHRR bands 1 and 2 is tracked at monthly
intervals for 8 classes (class numbers 7, 10-18, 27, and 46). Each class has its own territory in space and, depending
on the time of the year, has its own characteristic reflectivity. The
2-dimensional (2-d) space time spiral curves (ST-SCs) provide the best information on class behavior. For example, classes 27 and 46 experience big changes in a year
and have a large ?territory? in 2-d feature space (2-d FS). In contrast, classes
7 to 11 occupy narrow territories. Class 12-18 show moderate changes during the
year and a small territory in 2-d FS.
Space-time spiral curve (ST-SCs) to track class changes in 2-dimensional (2-d)
space and time.
approach is used to match and group classes that: (a) fall within similar 2-d
FS of a ST-SC plot, (b) have characteristic territory that leads to more
precise interpretation of the nature of the class (based on sound field
knowledge of at least one or more classes in a group). A large change in
territory implies agriculture and irrigation; very small changes imply forests,
plantations, continuous cropping; moderate changes imply rangelands. Time series analysis of NDVI and brightness temperature
Apart from the application of spectral similarity techniques
to group similar classes, we can also extract diagnostic information from the
vegetation dynamics shown by the time series itself, both independently and in
conjunction with crop calendars.
time series and cropping intensity
NDVI time series can categorize and identify irrigated area classes into
categories such as double crop, continuous crop, and single crop. Once the classes
have been identified using the approaches and methods described above, further
categorization is done using the time series NDVI plots. The example presented
here is for SPOT VGT data. Double crops exhibit swift rise in NDVI from 0.1 or
0.2 or less to 0.5 to 0.7 within 1-2 months and then quickly fall back to NDVI
of 0.2 or less. After 2-4 months of low NDVI, there is again a swift rise in
NDVI during the second crop. In contrast, continuous irrigated areas have NDVI
of about 0.3 or higher throughout the year.
the class identification process, time-series NDVI are plotted, compared, and
contrasted resulting in distinct categories. This is illustrated for 4 distinct
classes (a) dry shrub/grasses, b) rainfed crops, c) woody savannah and irrigated
crops in the Indus Basin, and d) for coniferous boreal forest) in Figure 19.
AVHRR NDVI spectral profile to identify and delineate classes.
During the class identification
process, the AVHRR time series earth skin temperatures were also plotted along
with time series NDVI (Figure 20). In the tropics, the greater the biomass
levels of a crop, the lower is the skin temperature and vice versa (Figure 20).
The skin temperatures of irrigated crops are low due to crop transpiration and
background moisture/wetness. In temperate climates, crops grown in summer
exhibit high NDVI and high skin temperatures. In contrast, during winter snow
and leaf-off conditions there is low NDVI and low skin temperature. Thus
the skin temperature time series helps identify LULC classes in different
climatic zones of the world.
AVHRR derived skin temperature versus AVHRR NDVI for semi-arid and tropical
5.5 Class refinement
Some classes from the
unsupervised classification could not be properly identified, as they appeared
to have characteristics of a number of other classes (which are anyway mixed
land-use categories). To resolve them and identify ones that have irrigation, a
number of decision tree models (e.g., Figure 21 and 22) were developed.
Further metrics had to be
incorporated to allow this refinement: 1) the principal components determined
for the SPOT and AVHRR time series, and 2) the locations of the classes were
investigated and matched against Forest Cover Density and GeoCover images for
both 1990 and 2000 at full resolution (15-30m) and using the ESRI 150-m
product. Typically the unresolved classes were split up into 10 to 20 sub-classes
before applying the decision tree and contextual ground truthing process.
5.5.1 Rule based decision trees
Figure 21 shows a rule based decision
tree to resolve one of the conflict classes, where annual rainfall is greater
than 2,400mm or 200mm per month (PGT2400). There were 4 conflict/mixed classes
(classes 13, 15, 18 and 19) which were masked, reclassified using SPOT monthly
time series NDVI, and labeled. The 4 mixed classes were resolved using
different approaches and discussed below.
21. Decision trees to resolve mixed classes.
Forest cover density (%) is used to resolve mixed class # 13 in the
precipitation > 2400mm per year segment.
Class13, labeled initially as ?Forest and some Irrigated?, was reclassified into five classes and then identified and
labeled based on existing LULC maps, GeoCover images and GT data. Based on the
forest cover density (FCD) the class 13 was resolved into 3 distinct classes
(Figure 21): cropland/woodlands mosaic, woodland/cropland mosaic, and evergreen
forest. This implies that the original class name ?forest and some irrigated?
was incorrect and further refinement of the classes showed no irrigation at
The decision trees for the
following mixed classes are not illustrated, but explain other types of class
Class15- Mixed: irrigated and savannas: SPOT NDVI was used to
separate irrigated area from savannas. However, results were not very satisfactory
due to persistent cloud problems over forest in the SPOT data. The same approach
adopted for Class 15 was used to the extract irrigated area class.
Class18- Irrigated (rice dominant) with some forest: the approach
used for Class 13 was repeated to refine this class.
Class19- Croplands with forest, grassland and some irrigated: This
class was further reclassified in to 10 clusters using all possible
combinations of SPOT NDVI, SPOT PCA, AVHRR NDVI (monthly 1999 and min), AVHRR
PCA, and tree cover density. Finally, we selected SPOT NDVI to enhance class
separability. Most of the classes separated well but subclasses 7, 8 and 9 were
a mosaic of forests and croplands and cloud cover. The cloud problem was
resolved using the SPOT NDVI MVC (1999) to reclassify them into five subclasses,
which were labeled based on the GeoCover 2000 images, secondary information and GT data.
17, in the PLT360-segment contains irrigated areas, natural vegetation and
grasslands. In order to resolve it, a decision tree algorithm was built (Figure
22). First, class 17 is masked out and reclassified using the time-series AVHRR
data into 10 sub-classes. Of the sub-classes, only class 6 remained mixed.
Class 6 is resolved using: (a) yearly NDVImax, NDVImin,
and (b) winter average NDVIs. When max NDVI>0.9 but min NDVI <0.1 it is only
irrigation. However, when max NDVI>0.9 but min NDVI>0.1, then it is a mix
of irrigation and grassland. The outcome is further division of class 6
into distinct categories of: (i) grassland and natural vegetation mix, (b)
irrigated areas (pure), and (c) irrigated areas mixed with grasslands.
22. Decision tree rules to resolve mixed classes.
The Maximum, minimum, and average NDVI were used, in a
decision tree framework, to separate out distinct areas within class 17.
5.5.2 Simple decision trees with Principal Components
Component Analysis (PCA) is also useful in resolving undefined classes. The
monthly AVHRR NDVI images over one year were stacked for class 20 in the PLT360-segment
and a PCA was performed. This resulted in 12-principal components (PCs). The month
of July had the greatest factor loading, accounting for most of the 99 percent
of variability explained by PC1. An unsupervised classification was performed
on the PC images resulting in 10 sub-classes. The decision tree helped resolve the
mixed classes. When (a) PC 1 (Jan) > 40, the class is irrigated; (b) PC 7, 8
and 9 (Jul, Aug, Sep) > 18 indicates mixed irrigation with woodland; and (c)
when PC 6, 7, 8 and 9 (June, July, Aug., Sept) < 18 it classifies as
When classes continue to be
mixed, in spite of the various methods and techniques discussed in previous
sections, we adopted the Geographical Information Systems (GIS) spatial
modeling approaches to resolve classes. This involved taking a mixed class and applying
GIS spatial modeling techniques such as overlay, matrix, recode, sieve and
proximity analysis. The GIS spatial data layers used include precipitation
zone, elevation zones, Koppen ecological zone, temperature zone, and tree cover
categories (see Figure 8). Any one or more of these classes help separate the
5.6 Class labeling
disaggregated classes from different segments of the World were combined to create
a single global disaggregated map of the World with 951 classes. The map, the
image, and associated product-line for the 951 disaggregated classes are
presented digitally in the IWMI?s Global irrigated area map (IWMI-GIAM) web site (http://www.iwmigiam.org). The
product is referred to as the ?Generic-IWMI-951-class-map?. Each of the 951 classes
also have class characteristics (Table 5) that include: elevation, tree cover,
precipitation, temperature, AVHRR 10-km temporal signatures from 1997-1998 (band
1 and band 2 reflectivity; band 4 and band 5 skin temperature, NDVI), and SPOT
1-km NDVI for every month during 1999.
We first labeled each of the IWMI-generic-951classes based on the classification techniques described earlier in this
section. Then we compared these with the corresponding global or regional
classifications (USGS LULC, USGS seasonal LULC, GLC2000, IGBP, and Olson eco-regions
of the world). The main objectives were to: (a) identify easy classes such as
forests and savannas and label them with a name that is consistent with
globally established class names, and (c) determine indicative class names that
will help further investigation to come up with a final class name. If, for a
tentative class, the assigned names from three or more global or regional
classifications match, the common class name is provisionally adopted. Table 6
illustrates the identification process for first twenty classes of the
AOAW-segment. The class naming is further verified using the hi-resolution
imagery, leading to a final class name. The process is repeated for each of the
951 classes and is very time consuming, but leads to robust results.
Table 5. Sample
characteristics of IWMI-951 class generic class map.
It is important to note that most
of the Global Product classes are quite consistent and it is straightforward
group classes such as forests, deserts and arrive at a consistent naming scheme
in the IWMI data. There are also other disaggregated global maps like the USGS
Seasonal LULC that has 253 classes (column 4, in Table 6) which, for example is
the only global classification that identified class 7 as irrigated. Therefore
we put greater emphasis on names arising from disaggregated classifications in choosing
a tentative class name.
The final class name can be quite
different if the classification resulting from our analysis and the match to high
resolution GeoCover or Google Earth imagery indicates the need for a better
Table 6. Indicative
class naming through the use of secondary data. Major global land use/land
cover (LULC) classifications were used in the class naming process. The main
purpose was to: (a) identify easy classes such as forests and savannas and
label them with a name that is consistent with globally understood class names,
and (b) determine provisional class names that will help derive a final class
|IWMI class # for Segment
Olson 1984 (96cl) (class name)
||USGS 1993 (17cl) (class name)
||USGS 1993 (255cl) (class name)
IGBP 1993 (17cl) (class name)
GLC 2000 (50cl) (class name)
||Inland water (Shallow)
||Water Bodies (Shallow)
||Water Bodies (Shallow),
||Water Bodies (Shallow)
||Water Bodies (Shallow)
||Inland water (Shallow)
||Water Bodies (Shallow)
||Water Bodies (Shallow)
||Water Bodies (Shallow)
||Water Bodies (Shallow)
||Inland water (Shallow), Rivers
||Water Bodies (Shallow)
||Water Bodies (Inland Water)
||Water Bodies (Inland Water)
||Water Bodies (Inland Water)
||Shallow water, Wetlands, Beaches
||Water Bodies, beaches,
||Water Bodies, beaches,
||Water Bodies, beaches,
||Water Bodies, beaches,
||Bare desert,Semi desert shrubs
||Barren or Sparsely Vegetated
||Barren and Sparsely Vegetated
||Barren or sparsely vegetated
||Sparse grassland, Bare rock, Stony desert
||Barren or Sparsely Vegetated
||Barren and Sparsely Vegetated
||Barren or sparsely vegetated
||Stony desert, Sandy desert and dunes
||Barren or Sparsely Vegetated
||Barren and Sparsely Vegetated
||Barren or sparsely vegetated
||Sandy desert and dunes, Stony desert
||Low sparse grassland, Crops and town, Crops, grass, shrubs
||Cropland/Grassland Mosaic, Shrub land
||Cropland (Rice, Wheat) with Woodlands, Grassland/Cropland (Small Grains)
||Cropland/natural vegetation mosaic, Croplands, Grasslands
||farmland, slope grassland
5.6.1 GeoCover, ground-truth and
The ESRI 150-m Geocover mosaic of
the world was extensively used in the class naming process. Unsupervised classes
from each segment were geo-linked with the 150-m GeoCover product to study the
class characteristics spatially. The value of Geocover is in: (a) providing
clear differentiation between agricultural lands and other land uses, (b)
obtaining spatial details of landscapes of interest; and (c) visualization of
patterns and features. These features enabled naming a class with greater
confidence and/or renaming a class with additional set of information. Some examples of the how GeoCover was used to identify and label classes is
illustrated in Figure 23.
Figure 23. Geocover Landsat
150-m data of the World in class identification and labeling process.
Examples of groundtruth from the Krishna and Ganges field campaigns are shown in Figure 24 (a-d) and examples of older IWMI ground-truth (see Figure 6) are shown in Figure 25 (a-d) for a broader range of locations
in the world. Digital photos (see Figure 7) sourced from the Degree Confluence
Project (DCP) are illustrated in Figure 26 (a-d).
Figure 24. Irrigated areas and
other LULC in the Ganges basin India.
Figure 25. Irrigated areas and LULC classes from different parts of
Irrigation in the Ganges includes
tube wells in alluvial areas, reservoirs, and river diversions
Figure 26. Irrigated areas and
LULC classes from different parts of the World from the degree confluence
The GIAM work lead to over 1000 classes. Over a two year
period, involving a core team of 8 members and 2 other support staff plus a
number of other useful contributors, these classes were resolved and labeled
using extensive interpretation techniques described in the previous sections
(see Figure 8). However, synthesizing these classes becomes extremely complex.
In order to avoid such a situation, we adopted a standard class naming
convention that involved the watering method, type of irrigation, crop type,
scale, intensity, location, and type of signature (see Figure 27).
5.7 Class aggregation and simplification
Generic-IWMI-951 disaggregated class map forms the basis of all other maps,
images, and associated product-lines. The products derived from the generic
28-class Global irrigated area map (GIAM10km-28 class);
8-class Global irrigated area map (GIAM10km-8 classes);
3-class Global irrigated area map (GIAM10km-3 class);
323-class Global Irrigated Area Map (GIAM10km-323 class);
229-class Global Map of Rainfed Cropped Areas (GMRCA229);
22-class map of Global Map of Rainfed Cropped Areas (GMRCA22);
76-class Global Map of Land use/Land cover Areas (GMLULCA75); and
10-class Global Map of Land use/Land cover Areas (GMLULCA10);
that had similar names and characteristics were grouped into a single class and
named uniquely. The process of aggregation is illustrated in Table 7.
GIAM10km-28 class irrigated area map is the main irrigated area product, but
two simplified GIAM10kmM-8 class, and GIAM10km-3 class maps have been produced
to ease visualization and understanding by irrigation practitioners. The GIAM10km-3
class map consists of the following classes:
- Irrigated, surface
- Irrigated, ground
27. Class naming convention. The standardized class naming convention is
depicted in this figure. At different levels,
the class naming may or may not
include a particular category such as the scale of irrigation or the intensity.
Table 7. Process of aggregation of classes from the generic map. The
irrigated area classes were aggregated from 951 class map
based on the methods discussed in sections 5,
6, and 7. Similar approach was used to aggregated classes into 28 or 8 or 3
Irrigated area classes from
Final Class names
Extracted from 951
recode to 102
recode to 53
recode to 34
273, 586, 622, 265
5, 7, 8, 12
Irrigated, large scale
Irrigated, in river valleys
Irrigated, in arid zones
near lakes and valleys
Irrigated, small scale
504, 498, 502, 499
10, 22, 23, 24
Irrigated, small scale,
mixed with rangelands
578, 215, 309, 500
16, 17, 18, 25
Irrigated, large scale,
double crop, mixed crops
Irrigated, single crop, more
Irrigated, single crop, less
Irrigated, deltas and
Irrigated, patches along
451, 492, 493
28, 29, 30
Irrigated, forest fragments
605, 501, 503
11, 49, 50
Irrigated, large scale,
double crop, rice dominant
47, 548, 506, 534,367,547,
46, 47, 62, 63, 43, 48, 39
31, 22, 20
Supplemental small scale
(pivot, drip), rainfed dominant
pivot dominant, grasslands
621, 366, 257
40, 42, 56
Supplemental, centre pivot
Supplemental, small scale
Supplemental patches, forest
468, 201, 469
54, 52, 53
581, 65, 623, 516
58, 59, 41, 61
26, 23, 27
LULC: irrigated fragments,
369, 467, 370, 36, 496, 562,
98, 99, 100, 86, 87, 88, 96
44, 46, 40
361, 362, 359, 48, 569, 360,
81, 82, 83, 73, 89, 92, 93
37, 39, 38
355, 354, 561, 545, 579,
563, 564, 565
85, 90, 91, 95, 84, 64, 65,
36, 48, 50, 33
irrigation, grasslands/shrub lands
491, 577, 528, 529, 476
71, 70, 68, 69, 72
45, 49, 35, 47
LULC: wetlands seasonally
482, 521, 63, 538, 619, 378,
348, 524, 344, 549
79, 80, 97, 101, 76, 102,
74, 78, 75, 94
41, 42, 51, 53, 52
LULC: forests, rainfed
croplands, irrigated mosaic
7 Estimation of irrigated areas using 3 methods
An estimate of the irrigated
areas of the world must take account of different crop seasons, cropping
patterns and intensity. In this analysis, we estimate area with respect to
three nominal cropping seasons (a) June-September,
(b) October-February, and (c) March-May. These represent the major
cropping seasons of the Indian sub-continent, and broadly cover similar seasons
in China, thus accounting for nearly half the world?s total equipped irrigated
area. For temperate areas with supplementally irrigated crops grown over one
long season, we avoid double counting across these the nominated seasons by
determining areas per pixel based on the time series signature. Since pixel
sizes are large at 1-km, and dominated by AVHRR time series at 10-km, it is
important to estimate the proportion of any one pixel that is irrigated in each
season. Use of total pixel area would result in a massive over-estimate. The full
pixel areas (FPAs) were converted to sub-pixel areas (SPAs) using irrigated
area fractions (IAFs). The overall procedure is shown in Figure 28. In order to
obtain reliable estimates of sub-pixel areas, we use 3 methods:
- Google Earth Estimates (GEE) (Figure 29);
- High resolution imagery (HRI) (Figure 30);
- Sub-pixel de-composition techniques (SPDT)
(Figure 31; and section 7.1)
The SPDT (Figure 31) and HRI
approaches provide irrigated area intensities for different crop growing
seasons (see Table 8), whereas the GEE approach provides net irrigated areas
28. Summary of area abstraction from the 28 irrigation class map.
Irrigated area fraction based on Google Earth Estimates
The IAF from Google earth
estimates (GEE) involves determining percent area irrigated for every GIAM10km-28
Class by zooming into Google earth images (e.g., Figure 36B). On an average at
least 30 points were randomly surveyed for every class and the irrigated area
fraction determined as the average area irrigated from all these points. The
process is repeated for all classes. The GEE approach acts as ?ground truth?
for the class.
Figure 29. Area estimation
using Google Earth (GEE).
For each GIAM10km-28 classes estimates of
irrigated area fraction (IAF) were made using Google Earth images.
points were taken for each class and averaged. The calculation for one class is
7.2 SPA of
pixels based on high-resolution imagery
second method of SPA estimation uses LandSat ETM+ images at 30 m resolution. At
least 3 hi-resolution images are downloaded per growing season for each of the
28 irrigation classes. The Landsat ETM+ grid is overlaid on the GIAM class and
images for estimation of the actual irrigated area within 10km pixels If a
class has 2 seasons, 6 images are downloaded and analyzed so that 3 images are
studied and averaged to determine the IAF in a given season.
Landsat images are first ?masked? to match areas defined in the global map (see
Figure 30). The image is then classified into 10 unsupervised classes. The
irrigated vs. non-irrigated areas are then identified using our class
identification schemes (see Figure 8). Then the IAF is the percent area
irrigated compared to total area of the masked Landsat image. Two other
methods were assessed (7.2.2 and 7.2.3), but were not as effective as this
Figure 30. Irrigated area
fraction from high resolution imagery (IAF-HRI). For each of the GIAM10km-28
The IAF-HRI were estimated by masking Landsat images for the area
occupied by the class and then determining irrigated vs.
images were also resampled to 10-km to match with AVHRR pixels and
co-registered (see De Fries, 1997). 325 AVHRR 10-km pixels are equivalent to
one Landsat image (185 x 170 km). The AVHRR NDVI from the 325 pixels are then
plotted against the Landsat ETM+ NDVI (?vegetation area fraction?) from the
resampled 10-km Landsat data. However, the resulting relationship was not clear
as a result of pixel size differences as well the problems associated with
precise co-registration. Hence the classification approach in section 7.3.1 is
Irrigated area fraction coefficient
times, a clear regression relationship between AVHRR NDVI and IAF with high R2-value
may be absent. In such a case, it will suffice to determine IAF for the entire
class based on the selected Landsat image by digitizing the irrigated vs. non
irrigated areas on the Landsat image. However, this approach is tedious and has
limitations of visual interpretation.
Sub-pixel decomposition technique
Determination of IAFs by
sub-pixel decomposition (SPDT) involves plotting AVHRR
band 1min (absorption maxima) versus AVHRR band 2max (reflection
maxima) of all the pixels in 10 sub-classes of a class and then scaling percentage
across them. The scaling is based on the knowledge base from ground-truth data,
digital photos, high-resolution images, literature, and relative positioning of
the pixels in the greenness-wetness-brightness areas in the RED versus NIR
the 28 irrigation classes is sub-divided into 10 giving a total class number of
280 for area estimation. The AVHRR band 1max and AVHRR band 2max
values for each sub-class are plotted, as for a BGW plot (e.g., Figure 31),
and the percentage area irrigated is determined based on the location of the
point in 2-d feature space (Figure 31). The percentage of irrigation is
assigned according to: (a) percent irrigated area canopy cover versus AVHRR
10-km band reflectivity and NDVI relationships from the Krishna and Ganges
groundtruth data; (b) percent cover recorded in IWMI Ground-truth data of the
World versus AVHRR 10-km NDVI or band reflectivity, and (c) extensive
literature review (Settle and Drake, 1993; Drake et al., 1997; Purevdorj et
al., 1998; Xiaoyang et al., 1998; Purevdorj and Tateishi, 2001; Barnes, 2000; Hallant,
2001 and Li et al., 2003). The actual irrigated area for a given class is
determined as the sum of the total pixel areas, multiplied by the sub-pixel
percentages for each of the 10 sub-classes.
Figure 31. Sub-pixel de-composition technique (SP-DCT)
the understanding one has of percent irrigated area versus band reflectivity, the
greater the reliability of the resulting area calculations. In this case, the
understanding comes from a combination of field and remote sensing experience
and is therefore limited by the geographical and farming system coverage
shows a detailed example of the plot for the first 20 full irrigation classes
and supporting plots for the other groups are provided in Annex 1, for further
reference and use by the reader. This has to be viewed in color to understand
the spatial position of classes and sub-classes.
de-composition plots are prepared for supplemental irrigated area classes: (a)
21, 23, and 24 and (b) classes 22, 25 to 28. Similarly two decomposition plots
were developed for LULC classes with some irrigation. The exact data and
assigned percentages in these plots can be progressively improved and expanded to different sub-groups of classes as the need arises. The decomposition plots are made
so that they can be easily modified at local, regional, national, and global
levels as new data becomes available. The relationship between AVHRR NDVI and sub-pixel
area percentage in the decomposition plots for the 20 irrigation classes is
presented in Figure 32.
Relationship between percent
irrigated area of classes 1-20
and AVHRR NDVI computed using band 1max
AVHRR band 2max
8 Accuracy assessment
A number of
different approaches were adopted to assess accuracies and errors (see
Congalton, 1994 and Foody, 2002). We concentrated on the irrigated area classes
and point based accuracy and error estimates were performed on two data sets
Irrigated Points classified as irrigated area
Accuracy of irrigated area
class = .................................................................................................... *
number of ground-truthed points for irrigated area class
Non-irrigated ground-truth points falling on irrigated
Errors of commission for
irrigated area = .................................................................................................... * 100
Total number of non-irrigated ground truth points
irrigated ground-truth points falling on non-irrigated
Errors of omission for
irrigated area = .................................................................................................... * 100
Total number of irrigated area ground truth points
Accuracy assessment makes use of three
distinct sources of reference data, so as to obtain a robust understanding of
the accuracies of the GIAM10km map V2.0 so that it can be compared to the Food
and Agricultural Organization and University of Frankfurt (FAO/UF) map of
global irrigated area. We also make a three way comparison for India, with reference to the Central Board for Irrigation and Power (CBIP). The distinct
sources of reference data are listed in section 8.1 to 8.3. The Google Earth
Estimates (GEE) data (section 8.3) are completely independent, randomly
generated. The degree confluence project (DCP) ground truth (GT) data (section
8.2) is relatively independent in that the DCP points are independent, but not
the other GT points. The other GT data (section 8.1) were also used in class
identification and labeling.
truth datasets from the Global Irrigated Area Mapping project
A total of
895 GT points were gathered by the GIAM project during 2004 and 2005 through a series
of groundtruth campaigns that included missions to all of India, separate missions to Krishna and Ganges basins, Sri Lanka, Uzbekistan, South Africa, and Mozambique. This data is far more refined for accuracy
assessment than the second dataset (section 8.2) because of its exclusive focus
on irrigated areas. However, we do not have broad coverage across the world.
A larger set
of ground-truth data with 1863 points is also used for accuracy assessment.
This dataset has far better spatial distribution across the world (e.g., Figure
7). However, the data itself comes from various sources that include: (a) Degree
Confluence Project (DCP), (b) various IWMI projects (e.g., wetlands, water
productivity), and (c) the GIAM project.
Accuracy assessments were also
made using 670 locations inspected in in Google earth at 30m pixel scale or
better. All GIAM irrigated area classes were combined into a single irrigated
class. The 670 sample locations were randomly chosen and their land use
determined in terms of: irrigated or not irrigated. These points were overlaid
on the irrigated area map and overall accuracy, errors of omission and errors
of commission were determined (Figure 32).
9.1 Global irrigated area map version 2.0 (GIAM10km V2.0)
The spatial distribution of the
irrigated area classes in the Global Irrigated Area Map (GIAM) are produced as
disaggregated map (GIAM10km-28 classes; Figure 33) and aggregated maps (GIAM10km-8
classes, Figure 34; GIAM10km-3 classes; Figure 35). GIAM10km-28 classes provides
information on irrigation type (surface water, ground water, and conjunctive
use), irrigation intensity (single, double, or continuous crop), and crop type.
The 8 class map provides watering method, irrigation type, and intensity. The 3
classes in the third map are: surface water irrigation, ground water
irrigation, and conjunctive use (surface and ground water) irrigation. The GIAM10km-28
class map has a complex set of classes and provides an understanding of their
distribution and class characteristics over time and space (Table 8).
The proportion of single, double
and continuous cropping allows calculation of areas based on cropping
intensities (i.e., single, double, continuous) leading to annualized areas
(summation of areas from different seasons). The cropping intensities and
calendars in Table 8 become more accurate if we look at individual countries or
sub-national administrative units.
Table 8. Characteristics of irrigated areas. Intensity
and cropping calendar for the GIAM classes in India.
The 8 class map provides single,
double, and continuous cropping for the surface water, ground water, and
conjunctive use irrigation. The 3 class (Figure 35) provides information on: (a) surface water irrigation; (b) ground water irrigation, and (c) conjunctive
(i.e., surface and ground water) use irrigation, without cropping intensity..
Figure 33. GIAM10km V2.0 28
Figure 34. GIAM10km V2.0 8 class map.
Figure 35. GIAM10km V2.0 3 class map.
9.2 Areas of irrigation
derived from GIAM10km map V2.0
The irrigated areas of the World
were estimated by the three methods (section 7) and the results are presented
First, the areas determined using
irrigated area fraction from Google Earth estimates (GEE) totals 412 million
hectares or Mha, without any specific information on cropping intensity.
The seasonal and annualized
irrigated areas are determined using irrigated area fraction from the high
resolution imagery and sub-pixel decomposition technique. For each of the 28
classes (Figure 33), we used the average IAF coefficients calculate seasonal
and annualized areas (summed over all seasons). The estimated total global
irrigated area for the 3 seasons are (Table 9a): (a) 263 Mha for season 1, (b)
176 Mha for season 2, (c) 41 Mha for season 3. The annualized global irrigated
area at the end of the last millennium was 480 Mha.
The areas have also been
summarized (Table 9b and 9c) for the 8 class map and the 3 class map.
The major finding of the IWMI analysis is that the net (412 Mha) and the annualized (480 Mha) cropped area under
irrigation very significantly exceeds the estimates of equipped area (274 million ha) by FAO, due to the extent of multiple cropping and private and
community developed irrigation. The area estimates in the map are derived for
each characteristic agricultural system around the world (e.g. long season
winter sown cereals in the northern hemisphere; triple rice cropping in SE
Asia; wet monsoon season (Kharif) and dry winter (rabi) systems in the Indian
sub-continent). The development of global irrigated area over last two
centuries is summarized in the Figure 36, with and without estimates of
cropping intensity. The presence of a large number of classes in GIAM10km-28 classes
(Figure 33) ensures varying seasonality of classes by taking more precise
cropping calendars between northern and southern hemispheres, the tropics, and
the higher latitudes. The aggregated maps (Figure 34 and 35 and Table 9b and 9c)
lose this distinction. The spatial characteristics of the GIAM class
information can be visualized using the higher resolution Landsat ETM+ re-sampled
150-m images, digital photographs, and google earth images from the specific
locations (Figure 37). The GIAM class information, presented in this manner is
of considerable value for the user who would like to have a ?visual picture?
Table 9a. Irrigated areas of the World from the GIAM10km-28
classes V2.0 map using IAF from HRI and SPDT.
Table 9b. Irrigated areas of the World derived from the GIAM10km-8
classes map V2.0.
Figure 36. Trends in irrigated
area since 1800. The IWMI estimate (http://www.iwmigmia.org) at the end of the last millennium
not only considered area irrigated, but also the intensity (i.e., area
irrigated during different seasons in 12-month period) and informal irrigation (e.g., ground water, tanks). This gives an estimate of 263 million hectares
during ?main? cropping season (Season 1) and a total of 480 million hectares
from 3 seasons: First Crop (263 Mha), Second Crop (176 Mha), and Continuous
crop (41 Mha).
Figure 37. The global irrigated area class snap-shot
illustrations for GIAM classes. The snap-shots (e.g., photos,
high-resolution images) of 4 distinct classes for a GIAM10km V2.0 class.
9.3 Irrigated areas of continents,
Countries, and river basins
Irrigated areas were also
calculated, based on combined IAF-HRI and IAF-SPDT, for the continents (Table 10a),
the Countries (Table 10b), and the IWMI and challenge program benchmark river
basins (Table 10c).
Of the 480 Mha annualized
irrigated areas in the world, 78 percent (375 Mha) is in Asia, 8 percent in
Europe, 7 percent in North America, 4 percent in South America, 2 percent in
Africa, and 2 percent in Australia. The area distribution for the seasons
follow similar trend (Table 10a). In Europe and North America overwhelming
proportion of irrigation is during the one main cropping season. In Asia, 154 Mha is irrigated in season 2 compared with 195 Mha during season 1. The surface
water irrigation in the world is 61 percent, the rest 39 percent is conjunctive
use (surface and ground water) and ground water. The surface water is well
separated. The ground water is often ingrained (and often dominates) in the
conjunctive use class.
Of the total global irrigated are
of 480,697,105 hectares, China (31.5 percent), India (27.5 percent) constitute
a total of 59 percent (Table 10b). The next countries have comparatively low
percentage irrigated areas: USA (5 percent), Russia (3.5 percent), Pakistan (3.3 percent). There are 9 Countries (Argentina, Australia, Thailand, Bangaldesh, Turkey, Kazakhstan, Myanmar, Uzbekistan, and Vietnam) with 1 to 2 percent. Brazil is ranked 15th with 0.85 percent (Table
10a). All other countries of the world have less than 1 percent or less
irrigated area. Forty countries have nearly 96 percent of all annualized irrigated
areas of the world (Table 10b). Normally (see Droogers, 2002, Postel, 1999), India is considered the leading irrigated area country, closely followed by China. However, our estimates show, China has 151 Mha of annualized irrigated area with India having 132 Mha. In the first season China with 76 Mha and India with 73 Mha are
close. However in the second season China has 68 Mha and India 54 Mha (Table 10b).
In summer there is only about 7 Mha in China and even less in India (about 6 Mha). The irrigated area fraction (IAF) for the classes in China were higher leading to greater sub-pixel area. For example, class 4 (see Figure 33)
which is mainly in China has IAFs of 0.53 and 0.67 for China. The class 8 and 24, 2 of the classes with substantial full pixel area (FPA), have
low IAFs. Class 8 for example, has IAFs of 0.37 for season 1 and season 2
bringing the sub-pixel area (SPA) down. Almost all previous irrigated area maps
either calculated areas based on FPA or from national statistics (which also
often ignores fallow areas).
The irrigated areas of the
continents and countries have been calculated based on the cropping calendars
and irrigated area fractions (IAFs) obtained from the global map. Our
expectation is that the calculation of irrigated areas for the countries will
be much more precise if cropping calendars are developed for individual
countries and irrigated area fractions developed separately for every country.
For this the GIAM team plans to work with national partners in 2007. However,
we do not expect the trends in irrigated areas to change and only a certain
(probably + 10 percent) adjustments to irrigated areas (maintaining the
present trend) is possible, especially for smaller countries.
Table 10a. Irrigated areas of the
Continents. The GIAM10km continental areas are compared with the FAO Aquastat
and the National statistics.
Table 10b. Irrigated areas of the
Countries. The GIAM10km country areas are compared with the FAO Aquastat and
the National statistics.
The irrigated areas of the IWMI
and CP benchmark river basins have been reported in Table 10c. Maintaining the
irrigated area trends of the continents and countries, the river basins of Asia and in particular India and China, dominate in irrigated area percentages (see Table 10c).
The annualized areas in Ganges is nearly 50 million hectares, Indus about 26
million hectares, and Yellow River about 20 million hectares. These are
staggering figures, given a basin like Nile with nearly 6000 years of
irrigation history has only about 5 million hectares.
Table 10c. Irrigated areas of the
river basins. The GIAM10km river basin areas are compared with the FAO Aquastat
and the National statistics.
9.4 Accuracy assessment of
the GIAM10km map V2.0 and its comparison with other maps
The accuracies were determined
through two methods:
- Ground-truth data
- Google earth data
First, we discuss accuracies
assessed using ground truth data and follow that with accuracies determined
using Google Earth data. Accuracies are assessed to determine whether the class
mapped is irrigated or not.
and errors of GIAM10km map V2.0 using ground truth data
There were 2 independent ground
truth data sets used in accuracy assessment. First a 895 point ground truth
data collected by GIAM team. Second the 1861 point ground truth data from the
degree confluence project (DCP).
Based on the GIAM teams 895
points, the accuracy of irrigated mapped as irrigated was 84 percent with 16
percent error of omission and 21 percent error of commission (Table 11a). In
comparison, the FAO map showed an accuracy of 79 percent with 21 percent for
errors of omission and commission (Table 11a). With DCP 1861 the accuracy
reduces to 77 percent but errors of omission and commission stay low at 23
percent. In comparison the FAO\UF V3.0 map shows an accuracy of 70 percent and
the errors of omissions and commissions of 30 percent (Table 11a).
and errors using Google Earth ground truth (GEGT) data for the World
The GEGT points are randomly
distributed around the world, with higher density of distribution of points
where irrigated area is dense. Accuracies using GEGT can be considered even
better than the ground truth data as a result of: (a) better distribution of
points around the world, and (b) precise spatial view of the landscape in
determining irrigation at 10-km scale which can often be unrealistic from the
The GEGT determined accuracy of
GIAM irrigated area classes to be 92 percent with very low error of omission of
8 percent and low error of commission of 17 percent (Table 11b). The FAO/UF
V3.0 map had an accuracy of 79 percent with higher errors of omission with 21
percent but lower errors of commission with 11 percent (Table 11b).
and errors for India in GIAM10km V2.0 for India
The accuracies and errors of the
irrigated area classes were also determined for India for two main reasons: (a)
the ground truth data for India is dense and well distributed as a result of
several GT missions, at various times, by the GIAM team; and (2) India is one
of the two largest irrigating nations in the world. Accuracies and errors are
determined for IWMI GIAM10km V2.0 India portion and compared with: (a) FAO/UF
V3.0 map, and (b) India?s Central Board of Irrigation and Power (CBIP) map.
The accuracy of GIAM V2.0 map in India was 86 percent with errors of omission of 14 percent and errors of commission of 20
percent. In comparison the FAO|UF V3.0 map and the India?s Central Board of
Irrigation and Power (CBIP) map have substantially low accuracies and higher
errors of omissions and commissions (Table 11a). In comparison, FAO/UF V3.0 map
had an accuracy of 76 percent with 24 percent errors of omissions and 26
percent errors of commission. The CBIP map had much lower accuracy at 61
percent and much higher errors of omission (39 percent) and errors of
commissions (23 percent). This is because, the CBIP irrigated area map for
India almost completely ignores groundwater irrigation, conjunctive (surface
water plus ground water) use within irrigated areas, and the supplemental
irrigated area as its focus is almost completely on large scale surface water
irrigated areas with some medium to small scale surface water irrigated areas.
The trends in accuracies and errors between GIAM V2.0,
FAO\UF V3.0, and CBIP using the Google Earth ground truth (GEGT) remains the
same, with higher accuracies and lower errors in GIAM V2.0 (see Table 11b).
9.5 Accuracy assessment discussions
Overall, the results show that
the accuracies of the IWMI GIAM V2.0 was about 7 to 12 percent higher than FAO\UF
V3.0. The errors of omission and commission were only slightly better in GIAM.
The area calculations in the two maps differ significantly since IWMI GIAM uses:
(a) intensity of irrigation to obtain irrigated areas based on seasons, and (b)
sub-pixel decomposition techniques to obtain the irrigated fraction within a
pixel. The FAO/UF relays on the country statistics that may not include
intensity. Further, there is no direct link between FAO spatial data and area
statistics. The areas are reported directly from country statistics and the
spatial distribution of irrigation is ?adjusted? to fit the country statistics,
using known extents of surface irrigation and other secondary information. The India?s CBIP under-estimates irrigation since it, largely, ignores informal (e.g., ground water) irrigation. For India portion GIAM10km V2.0 map accuracies and
errors were significantly better than that of FAO/UF V3.0 and CBIP (Table 11a and
11b). Especially, the errors of omissions and commissions were much better,
indicating that IWMI GIAM is picking the informal (e.g., small reservoirs,
tanks, ground water) irrigation better.
There are fundamental issues
related to accuracy assessments at such large scales as 1-km or 10-kilometer
resolution pixel size. There are considerable difficulties in ground truthing
and establishing the exact percent of area irrigated in a 1-km x 1-km (100
hectares) and especially at 10 km x 10 km (or 10,000 hectares) resolutions. For
example, when GT data is collected in a portion of a pixel that has land cover
other than irrigation and has irrigation in patches (say 25 percent of pixel
area), we may not even see irrigated portions during GT data collection. This
will lead to the pixel being labeled ?other LULC? in GT data, but whereas in
reality it has 25 % irrigation. Satellite sensors capture the average
reflectivity from the pixel and hence are influenced by both the irrigated as
well as non-irrigated components within the pixel leading to a average spectra
for the pixel. Whereas satellite data distinctly shows the difference in a
pixel with zero irrigation and one with 25 percent irrigation, GT data often
fails to do so. This will lead to situations such as, for example: (a) rainfed
GT points falling on a pixel mapped as irrigated (commission error); (b)
irrigated GT points falling on a pixel mapped as other LULC (omission error). This
can lead to somewhat higher omission and commission errors. The phenomenon is
acute when dealing with pixels of low percent (<20) of irrigation which have
greater likelihood of being labeled as classes other than irrigation, resulting
in highly exaggerated commission errors. This also implies an area based
accuracy assessment maybe more powerful and robust than point based accuracy
assessment. However, quality area based reference data is nearly non-existent.
Offset against this spatial advantage of remote sensing, is the fact that there
are multiple reasons for an average pixel scale signal, and it is therefore
possible to confound on interpretation with another reality. The very high
resolution (sub-meter to 4 meter) images available in google earth facilitate
determining the land cover and irrigation structural patterns which will be
invaluable in determining irrigation vs. non-irrigation. Hence, the GEGT is
considered a better data for accuracy assessment.
The accuracy assessment comparison
between the GIAM10km, FAO/UF, and CBIP maps (Table 11a and 11b) are indicative
and not definitive. In a strict sense, none of these maps can be directly
compared with each other as a result of considerable differences in scale/resolution,
primary datasets used to derive the map information, and due to differences in methods,
techniques, and approaches on how these maps are derived.
Even if the accuracies and errors
between the IWMI GIAM10km and FAO/UF maps are similar, the calculated areas
differ as a result of fundamental differences in how the maps are produced. In
IWMI-GIAM, the global annualized (i.e., taking cropping intensity or seasons)
irrigated area is 494.4 Mha and the net areas per season are: 278 Mha, 176.5
Mha, and 39.9 Mha (Figure 33 and Table 9a). In contrast the other global
irrigated area estimates 274 Ma (Siebert,et al., 2005; Siebert, S., Döll, P.,
Hoogeveen, J., 2002), in which the FAO/UF provides area ?equipped for
irrigation? to be 274 Mha (Siebert,et al., 2005; Siebert, S., Döll, P., Hoogeveen, J., 2002).
Accuracies can be similar, but areas can differ because:
- Intensity (seasonality)
consideration: The IWMI GIAM10km V2.0 provides gross areas based on
cropping intensity (single crop, double crop, triple crop, continuous
crop). Other area estimates count the area once (net) based on area equipped and assuming irrigation once during a major cropping season;
- Sub-pixel fraction differences:
The irrigation fraction in IWMI method depends on the 3 methods (GEE, HRI,
and SPDT). The FAI/UF is dependant on Country statistics at sub national
- Area estimation approaches:
The FAO/UF area calculations are dependant on the national statistics and
their extrapolation onto spatial maps. The IWMI GIAM is interpreted
directly from the satellite image characteristics.
Table 11a. Accuracy assessment of IWMI GIAM V2.0 Vs.
FAO/UF V3.0 vs. CBIP using ground truth data. The IWMI Global irrigated area
map (GIAM) is compared with the: (a) Global irrigated area map of the
FAO/Frankfurt University, and (b) India's Central Board of Irrigation and Power
Table 11b. Accuracy assessment of IWMI GIAM V2.0 Vs.
FAO/UF V3.0 vs. CBIP using Google Earth ground truth (GEGT). The IWMI Global
irrigated area map (GIAM) is compared with the: (a) Global irrigated area map
of the FAO/Frankfurt University, and (b) India's Central Board of Irrigation
and Power (CBIP).
A discussion on mapping irrigated areas and comparison of maps
area maps of IWMI GIAM10km V2.0, Food and Agricultural Organization\University
of Frankfurt (FAO\UF) V3.0, and India?s Central Board of Irrigation and
Power (CBIP) maps are compared and discussed. We shall begin with detailed
illustrations of comparisons in India where we have detailed ground truth data
collected by the GIAM team and very reliable and detailed maps from the CBIP. The
extensive ground truth data collected during the field campaigns were
invaluable in these comparisons.
10.1 Major irrigation
First we shall illustrate a
comparison of maps for major irrigation. The CBIP map, basically, represents
major irrigated areas (leaving out informal irrigation) and is considered
accurate for major command area irrigation. For the purpose of comparison we
took 5 random ground truth (GT) points falling within CBIP map (Figure 38b) and
overlaid them on the IWMI GIAM10km V2.0 for India (Figure 38a) and CBIP irrigated
area map for India (Figure 38b). According to GT data, 2 were informal (tank, ground water) irrigation, 1 was major irrigation, 1 naturally irrigated, and 1
rainfed. The GIAM10km classes (Figure 38a) showed 3 informal (2 conjunctive and
1 ground water) and 2 surface water. The CBIP (Figure 38b) showed all points as
major irrigation. These results clearly implied that the GIAM10km has closer
match with ground reality in terms of type of irrigation.
10.2 Informal irrigation
Next, we illustrate how well the informal (e.g., small reservoirs, tanks, ground water) irrigation is captured between maps. For
the purpose, we randomly select 5 GT points with informal irrigation. The CBIP
map misses all the randomly selected groundwater check points (Figure 38c). The
GIAM10km V2.0 idenitifies all of them- 3 as conjunctive and 1 as ground water
irrigation (Figure 38c). This is very close to ground truth data which also had
3 supplemental irrigated area classes. Finally, identification of small scale
irrigation from minor reservoirs and ground water is illustrated in Figure 38e
for CBIP and Figure 38f for IWMI GIAM10km V2.0. Of the 5 randomly chosen GT
points (2 irrigated small scale, 1 irrigated large scale and 2 rainfed) CBIP
misses all (Figure 38e) whereas GIAM maps 3 as conunctive and 2 outside. It
actually maps 2 correctly as informal irrigation, 1 rainfed correctly as
?outside? irrigated areas. Of the other two points, it maps a rainfed class as informal irrigation and and informal irrigation as ?outside irrigated area?. Leading to some
omission and commission errors. However, as we have seen in Figure 38a through
44f, informal irrigation is well captured in GIAM10km.
Figure 38. Evaluation of the GIAM
for large scale, small scale, informal, and supplemental irrigation. The
IWMI GIAM and India?s Central board of irrigation and power (CBIP) irrigated
area maps are evaluated for: (a) large scale irrigation- (Figure 38a and 38b);
(b) informal irrigation such as ground water and tanks (Figure 38c and 38d);
and (c) small scale (e.g., minor reservoirs) irrigation (Figure 38e and 38f).
10.3 Comparing global products
The comparison between FAO/UF
global irrigated area map and GIAM10km V2.0 highlights the distinct features of
areas where the two maps: (a) perfectly match (e.g., Figure 39a in Upper Ganges basin), (b) broadly match (e.g., Figure 39b in Cauvery delta), and (c) do not
match at all (e.g., Figure 39c in Ganges delta). This illustration is a
?representative? comparison of the two global irrigated area maps as we see the
similar trends in other places of the World. The countrywise area statistics of
the 40 best ranked irrigated area Countries of the World are plotted for the
two maps taking FAO\UF Country statistics and IWMI GIAM10km V2.0 season 1 areas
which showed an R2 value of 0.92 (Figure 40).
Figure 39 (a-c). Comparison of the two global irrigated
area maps: GIAM10km V2.0 and FAO/UF V3.0.
Figure 40. Comparison of irrigated areas of 40 leading
countries between IWMI GIAM10km V2.0 vs. FAO\UF V3.0.
11.0 Irrigated area class
At this stage it is useful, to discuss the
issues involved in final class labeling and the approach used in IWMI GIAM10km
V2.0. Classes were named based on a set protocol and rigorous methods (see
Figure 8) that had a clear class naming convention (Figure 27). In addition, the
final class labeling (see Figure 33, 34, and 35) was also based on consultation
with irrigation experts so that the class names represent the commonly
understood meaning of a particular irrigation type. More generic and detailed
names are provided in GIAM10km-28 classes (Figure 33) and a much simpler,
broadly understood names are provided in GIAM10km-8 classes (Figure 34), and GIAM10km-3
classes (Figure 35).
The final class labeling were categorized under
the following groups:
water, single crop, crop type or dominance.
water, double crop, crop type or dominance.
water, continuous crop, crop type or dominance.
The watering method (irrigated or rainfed) and
irrigation type (surface, ground or conjunctive use) are determined based on
the protocols and methods (see Figure 8 and sections 5 and 6). The single,
double or continuous crop is determined based on the spectral signature for
every class based on time-series satellite imagery (see example in for class 1
and 4 in Figure 41). The same class 1 and 4 also occur in Iran showing somewhat different signature characteristics (in magnitude and timming of
peaks and lows). Indeed, it is possible to get a cropping calendar for every
pixel of irrigated area classes by simply clicking on any point on irrigated
area class and looking through the time-series imagery of a mega-file as we
have done in Figure 41 and 42. Final variable (crop type or dominance) in
naming is based on ground truth data and literature.
The above naming convention is repeated for
ground water and conjunctive use irrigation.
Figure 41. Single
crop (red) and double crop (cyan) irrigation in lower Ganges.
Figure 42. Double crop (left) and single crop (right)
irrigation in Zahandeh and Rud.
In Figure 33, Classes 1-10 are surface water
irrigation, classes 11-15 are ground water irrigation, and classes 16-28 are
conjunctive use irrigation. These classes were combined appropriately to
produce the simplified 8 call and 3 class irrigated area maps (see Figure 34 and
The class labeling process of one class has been
discussed in detail. First, we go through a normal protocol (Figure 8), Methods
(sections 5 and 6), and class naming convention (Figure 27). In addition, the
detailed approach to name a class is illustrated below for one class. The class
28 (Figure 43) was labeled ?irrigated, conjunctive use, continuous crop, mixed
crop? in GIAM10km - 28 class map (Figure 33). It occurs mainly in the Pampas of
Argentina, which is predominantely rainfed. However, different degrees of
supplemental groundwater irrigation (e.g., pivots, drip) and some pumping from
rivers also exist. Centre pivot irrigation is used in humid plains of pampas to
supplement rainfall (Maletta, 1998 and Maletta, 1999).
The spectral characteristics of the class show
near continuous cropping, with AVHRR NDVI greater than 0.35 or more throughout
the year (Figure 43). Rainfall during May-September is low, averaging less than
40 mm per month (see Figure 43) and is insufficient to sustain such high
vegetation in an agricultural belt. The pampas is a humid plain, which is very
flat and is poorly drained. This and man made obstacles such as roads and
railroad embankments lead to flooding and waterlogging for months, favoring
growth of weeds and natural pasture in the vicinity even during relatively dry
spells (Maletta, personnel communication). The long period of deficit rainfall,
and continuously high NDVI strongly implies some degree of irrigation.
Evidence (Maletta, personal
communication) from the field suggest that centre-pivot irrigation in the Pampas is mostly used for complementary and drip irrigation is used for horticulture.
Maletta summarizes the situation: ?There is indeed a need to irrigate more, as
witnessed by the fact that average yields (especially for maize) are quite
below potential. But (1) massive use of irrigation is not yet happening, (2)
aquifers may not support such an extensive use of underground water, and (3)
gravity irrigation is in general difficult due to very flat land, thus
requiring pumping (which is not generally done) from the many streams flowing
through the plains.? The data from the Government administration (http://www.indec.mecon.ar/) shows nearly
1.4 million hectares irrigated. These do not account for an occasional
irrigation (e.g., one or two irrigations during the cropping period, during
deficit rainfall periods) or informal (individual farmers irrigating without
governmental knowledge mainly through ground water pumping). Overall, the pampas
region depends on rainfall, but has significant proportion of irrigated land
(pivots, drip, river pumps), humid flat-water logged regions and scattered
informal irrigation. These characteristics lead the class to be named: ?conjunctive
use?. In the past, irrigated area maps only included areas with formal canal
networks and major works such as reservoirs or barrages. But many parts of the
world have various levels of irrigation that need to be accounted to obtain a
realistic estimate of actual irrigated areas.
Figure 43. Evaluation of GIAM for conjunctive irrigation.
The rainfed class with significant central pivot supplemental irrigation in the
Pampas in Argentina.
V2.0 products and dissemination
The IWMI GIAM10km V2.0 data and
products are distributed via dedicated web page at:
The web page consists of GIAM10km
V2.0 products at global level mapped at 1-10 kilometers and include maps, images, class characteristics, area calculations,
snap-shots (high-resolution images) and photos, animations, and accuracies. The
products are made available at nominal resolution of approximately 1 kilometer
since all data were resampled and analyzed at 1-km scale. However, we urge the
users to treat it as a nominal 10 km2 since a overwhelming
proportion of the data used in analysis is at this scale. But it must be noted
that a significant proportion of the mega data used in analysis include SPOT
time-series for 1999 and GTOPO30 were at 1-km. GIAM10km map is also available
for Google Earth, please download GIAMv2.kmz file from the home
page of GIAM main site (http://www.iwmigiam.org)
primary GIAM10km V2.0 products are:
GIAM10km V2.0 28 class map (GIAM10km-28 classes);
GIAM10km V2.0 8 class map (GIAM10km-8 classes);
GIAM10km V2.0 3 class map (GIAM10km-3 classes).
The website additionally contains
three other global agriculture products and their associated documentation:
Global map of Rainfed Cropped Areas (GMRCA)
229 class map
- Aggregated 22
Global map of all land use/land cover (LULC) areas (GMLULCA)
76 class map
- Aggregated 10 class map
Global IWMI generic 951 class map (Generic-IWMI-951)
International Water Management Institute (IWMI) has produced a global irrigated
area map at 10 kilometer scale (GIAM10km V2.0), for the end of last millennium,
using remote sensing data. The total annualized irrigated areas of the World
are 480 million hectares (or Mha). Globally, the area available for irrigation
is 412 Mha. Annualized area takes into consideration irrigated areas during
different seasons over same areas within a given year. Of the total annualized
area of 480 Mha, a total of 75 percent (375 million hectares) of all irrigated
areas of the world is in Asia, followed by Europe with 8 percent North America with 7 percent, South America 4 percent, Africa 2 percent, and Australia 2 percent. The irrigated areas spread across the season are: (a) 263 Mha for
season 1, (b) 176 Mha for season 2, and (c) 41 Mha for continuous.
Two Countries, China and India, together have a staggering 59 percent (284 Mha) of all the Global annualized
irrigated areas. Of the 59 percent, China has 31.5 percent and India 27.5 percent. China has an annualized area of 151 million hectares and India 132 million hectares. The first or the major cropping seasonal areas follow similar
pattern to annualized areas. China and India have extensive double cropping. In
the first season China has 76 Mha irrigated, followed by 68 Mha in the second
season. In India, the area irrigated is 73 Mha in first season and 54 Mha in
second season. The next leading irrigated area countries (as a percentage of
the global annualized sum of 412 Mha) are USA (5 percent), Russia (3.5 percent), and Pakistan (3.3 percent). There are 9 countries (Argentina, Australia, Thailand, Bangladesh, Turkey, Kazakhstan, Myanmar, Uzbekistan, and Vietnam) between 1 to 2 percent. Every other country in the world,
individually, only has less than 1 percent area irrigated. The 40 leading
irrigated area countries have nearly 96 percent of all irrigation in the World.
Surface water irrigation is 61 percent and the rest (39 percent) is conjunctive
(surface and ground water) or pure ground water.
There are three global irrigated
area maps produced by GIAM team: GIAM10km 28 class map, GIAM10km 8 class map,
and GIAM10km 3 class map. The classes represent: (a) irrigation by surface
water, ground water, and conjunctive use; (b) cropping intensity (e.g., single
crop, double crop, and continuous crop) are provided for every class; and (c)
crop type or dominance. The accuracy of mapping irrigated areas were determined
using 3 independent datasets- 2 ground truth data sets and 1 Google earth
estimate dataset. The accuracies varied between 84 to 91 percent, the errors of
omissions less than 16 percent, and errors of commission less than 21 percent.
The results of our study were compared with the irrigated area map statistics
of the Food and Agricultural Organization of the United Nations (FAO) and the
University of Frankfurt (UF) version 3.0 (FAO\UF V3.0). The FAO\UF used
National statistics and GIS techniques to derive irrigated areas. FAO\UF V3.0
determined ?area equipped for irrigation? (but not necessarily irrigated) for
the World as 271 Mha which is quite different from GIAM10 km V2.0 TAAI of 412 Mha. For the leading 40 countries, which constitute 91 percent of all
irrigated areas of the World, the GIAM10km V2.0 season 1 areas versus FAO\UF
V3.0 areas had a R2- value of 0.90.
The key achievements of the GIAM10km
V2.0 work have been:
- Methodology development:
a comprehensive set of methods and techniques for mapping irrigated areas
of the World using remote sensing data at various scales or pixel
resolution has been developed (see this research report and also earlier
work by Thenkabail, et al., 2005, Thenkabail et al., 2006, Biggs et al.,
in approaches and datasets: mega-file compositions through fusion of
multi-resolution time-series imagery;
in methods: hyperspectral techniques for multispectral time-series
mega-file imagery. The methods include spectral matching techniques (SMTs) and
space-time spiral curves;
identification and labeling: Rigorous strategies for class identification
and labeling have been developed. Strategies for resolving mixed classes
through GIS modeling in which wide array of secondary datasets have been used
has been established;
areas (SPAs) and irrigated area fractions (IAFs): Innovative sub-pixel area
(SPA) calculation methods using irrigated area fractions (IAF) has been
developed. Three IAF methods were developed: (a) IAF based on Google earth
estimate (GEE), (b) IAF based on high resolution imagery (HRI), and (c) IAF
based on sub-pixel decomposition technique.
areas calculated from remote sensing are, almost always, reported as full pixel
areas (FPAs). But the correct areas can be only obtained through SPA. This is
especially true in coarser resolution imagery. Development of practical methods
to obtain SPAs through IAFs is, thereby, a highly significant achievement.
- Annualized areas (or
Intensity of irrigation) and irrigated area fractions (IAFs): The
study determined and provided IAFs through 3 methods. The irrigated area
fractions from the Google eye (IAF-GEE) when used to multiply the full
pixel areas (FPAs) provide total area available for irrigation (TAAI). The
IAF from high resolution imagery (IAF-HRI) and sub-pixel decomposition
technique (IAF-SPDT) can be obtained for different seasons (e.g., season 1
crops, season 2 crops and so on). The seasonal IAF coefficients helped
determine irrigated areas of every class for season 1, season 2, and
continuous. Annualized (or intensity) is summation of season 1, season 2,
and continuous. The coefficients of IAF-HRI and IAF-SPDC were combined to
provide more robust SPAs.
areas are very unique. Ability to determine annualized areas has huge
implications of the intensity of irrigation in given land and the implications
in determining the quantum of food production and water consumption.
- Informal irrigation: The
GIAM10km demonstrated the ability to map informal irrigation (i.e.,
irrigation from minor reservoirs, tanks, and ground water) well. This is
especially crucial given the quantum of informal irrigation in the world,
especially from millions of tube wells.
- Crop characteristics:
Every class (or for that matter every pixel within a class) will have its
own characteristics in terms of its vegetation dynamics and seasonality. GIAM10km
product is not just a map. It is a dynamic tool from which one can study
variables such as cropping calendars, crop growth stages, biomass levels,
and fraction areas irrigated.
- Precise location of
irrigated areas: Most irrigated area maps provide areas without
showing precise location of irrigated areas. For example, an entire state
or country is often shown to have certain percentage area irrigated
without showing where exactly it is. GIAM10km map provides precise
location. The errors of omissions (less than 16 percent) and commissions
(less than 21 percent).
- Product line: GIAM data
and products are made accessible online for free as a global public good
(GPG) from anywhere in the world (http://www.iwmigiam.org).
The products consist of, for example, irrigated area maps, statistics,
20-year every month animations, snap-shots of higher resolution imagery to
help visualization of classes, class characteristics, irrigated area
fractions for area calculations, methods, and datasets.
Within the scope of GIAM project,
irrigated areas are also mapped at 500-m resolution for India, as a start, and 30-m resolution for the Ruhuna basin in Sri Lanka and Krishna basin in India.
Currently, IWMI is in the process
of developing a joint vision and strategy with FAO\UF on global irrigated area
mapping. We are also developing partnerships and collaborations with the
National Governments and Institutes. To that end, work continues on the
development of techniques to map and test the accuracy of classification across
the full extent of the Indian sub-continent (Pakistan, India, Bangladesh) and
Sri Lanka, covering a range of agro-ecologies and degrees of difficulty for
remote sensing (in terms of cloud cover, heterogeneity and scale of landscapes
and land use). The work is expected to be expanded to China and other Countries. A Consortium for Irrigated Area Mapping and Assessment (CIAMA)
is expected to be set up, with an array of International partners, during the
GIAM2006 International Workshop to be held in Colombo, Sri Lanka.
The team seeks feedback from all
users, readers and interested parties, and continues to harvest ground-truth
data to verify and upgrade the map. The team welcomes any feedback on the
methods and results, and actively seeks to expand the available ground-truth in
order to build a global ground-truth database within the IWMIDSP (http://www.iwmidsp.org ). All the imagery
and documentation associated with GIAM are made available through the dedicated
The products consist of maps,
images, class characteristics, area calculations, snap-shots, animations, and
accuracies. It is our hope that these products will, in time, be a useful
resource for the remote sensing and water management community ? both for
researchers and practitioners.
I.P. and Yadav, Jai Singh Pal and Massoud, F.I., 1988. Salt Effected soils and
Their Management.. Food and Agriculture Organization of the United Nations, Rome.
S., Joshi P.K., Shukla Y. and Roy P.S. 2004. Spot-Vegetation multi temporal
data for classifying vegetation in South Central Asia. Current Science 84(11):
Bartholome“, E., and
BELWARD, A.S. 2005.
GLC2000: a new approach to global land cover mapping from Earth observation
Journal of Remote Sensing, 26(9):1959-1977.
E.M., Clarke,T.R. and Richards, S.E.. 2000. Proceedings of the Fifth
International Conference on Precision Agriculture.ASA-CSSA-SSSA, 677 South Segoe Road, Madison, WI 53711,USA.
Thenkabail, P., Krishna, M.G., Scott, C., Parthasaradhi, G.R., Turral, H. 2006.
Vegetation phenology and irrigated area mapping using combined MODIS
time-series, ground surveys, and agricultural census data in Krishna River
Basin, India. International Journal of Remote Sensing. (in review).
Campbell, J.B., 1987. Introduction to Remote Sensing. The Guilford Press, New York. pp. 551.
R. Hansen, M., Townshend, J.R.G., Janetos, A.C., and Loveland, T.R.. 2000a.
Continuous Fields 1 Km Tree Cover. College Park, Maryland: The Global Land
R. Hansen, M., Townshend, J.R.G., Janetos, A.C., and Loveland, T.R. 2000b. A
new global 1km data set of percent tree cover derived from remote sensing.
Global Change Biology. 6: 247-255.
R. S., and Townshend, J. R. G., 1994, Global land cover: comparison of
groundbased data sets to classi. cations with AVHRR data. In Environmental
Remote Sensing from Regional to Global Scales, edited by G. M. Foody and P. J.
Curran (New York: Wiley), pp. 84?110.
R., Hansen, M., and Townshend, J., 1995, Global discrimination of land cover
types from metrics derived from AVHRR Path. nder data. Remote Sensing of
Environment, 54, 209?222.
R., Hansen, M., Townsend, J. G. R., & Sohlberg, R. (1998).Global land cover
classifications at 8 km resolution: the use of training data derived from
Landsat imagery in decision tree classifiers. International Journal of Remote
Sensing, 19, 3141? 3168.
Siebert, S., 1999. A digital global irrigated area map. Kassel World Water
Series 1, Center for Environmental Systems Research, University of Kassel, Germany, 23 pp + Appendix. [Download pdf,]. Version 2: Update for Latin America and Europe; new methodology to combine statistical and geographical information
Siebert, S., 2000. A digital global irrigated area map. ICID Journal, 49(2),
N.A., Mackin, S. & Settle, J.J. 1997. Spectral matching and mixture
modelling of SWIR AVIRIS imagery, in Proceedings of the 23rd Annual Conference
of the Remote Sensing Society: RSS?97 Observations & Interactions, Reading, 410-415
P., 2002. Global irrigated area mapping: Overview and recommendations. Working
Paper 36. International Water Management Institute, Colombo, Sri Lanka. pp. 54.
W.H. and Harsanyi J.C., 1997, ?Mapping the distribution of mine tailings in the
Coeur d?Alene River Valley, Idaho, through the use of a constrained energy
minimization technique?, Remote Sensing of Environment, No59, pp 64-76.
J., D. F. Heath, K. F. Klenk, N. Oslik, K. D. Lee, H. Park, P. K. Bartia, and
D. Gordon. 1984. User's Guide for the Solar Backscattered Ultraviolet (SBUV)
and the Total Ozone Mapping Spectrometer (TOMS) RUT-S and RUT-T Data Sets: October 31, 1978 to November 1980. NASA Reference Publication 1112.
M., Boyd, D. S., and Curran, P. J., 1996, Relations between tropical forest
biophysical properties and data acquired in AVHRR channels 1± 5. International
Journal of Remote Sensing, 17, 1341± 1355.
Granahan J.C. and Sweet J.N, 2001, ?An Evaluation Of Atmospheric Correction Techniques Using The
Spectral Similarity Scale?, IEEE 2001 International Geoscience and Remote
Sensing Symposium, 2001. Vol. 5, pp. 2022-2025.
L. H. J., Ian, P.P., Chris, M. and Priestley G. 2001. Prediction od sheet and
rill erosion over the Australian Continent, Incorporating monthly soil loss
distribution. Technical Report 13/01, CSIRO Land and water, Canberra, Australia
and Hilmy Sally. 2001. Linkage between Productivity and Equitable Allocation of
Water, 2nd WARSFA / WaterNet Symposium: Integrated Water Resources Management:
S. and Roux M., 2003, ?Material Mapping from Hyperspectral Images using
Spectral Matching in Urban Area?, Submitted to IEEE Workshop in honour of Prof.
Landgrebe, Washington DC. USA Oct. 2004.
1990. International Geosphere-Biosphere Programme, 1990. Global Change:
Report No. 12, Stockholm, Sweden.
K., 1991. NOAA Polar Orbiter Data User's Guide. NCDC/SDSD. National Climatic Data Center, Washington, DC.
F.N., and Zhu, X. 2001. Evolution of long-term errors in NDVI time series:
1985-1999. Advances in Space Research. 28 (1): 149-154.
Li, X. B.,
Chen Y. H., Shi P. J. and Chen J, 2003. Detecting Vegetation Fractional
Coverage of Typical Steppe in Northern China Based on Multi-scale Remotely
Sensed Data. Acta Botanica Sinica 2003, 45 (10): 1146-1156
Loveland, T.R., Reed, B.C., Brown, J.F., Ohlen, D.O., Zhu, J,
Yang, L., and Merchant, J.W., 2000. Development of a Global Land Cover
Characteristics Database and IGBP DISCover from 1-km AVHRR Data. International
Journal of Remote Sensing, v. 21, no. 6/7, p. 1,303-1,330.
Maletta, H. 1998.
Irrigation in Brazil: Overview and policies. Inter-American Development Bank - NOMISMA Spa, Brazil - Agricultural Sector Review, Bologna (Italy), 1998.
Maletta, H. 1999. Thematic study: Irrigation -
Latin American paper - Draft 1 - Date: 31/10/1999. World commission on Dams.
S. 2002. Water for Agriculture: irrigation agriculture in International
perspectives. Spon?s Environmental Science and Engineering Series. Pp. 235. Taylor and Francis, New York, NY, USA.
T.D., Carter, T.R., Jones, P.D., Hulme,M., New, M., 2003: A comprehensive set
of high-resolution grids of monthly climate for Europe and the globe: the
observed record (1901-2000) and 16 scenarios (2001-2100). Journal of Climate:
Kaleemulah,Zaigham Habib,Saim Muhammad. 2001. Spatial Distribution of
References and Potentia Evapotranspiration,IWMI Working Paper 24,Pakistan Country Series Number 8
1994. 5 Minute Gridded World Elevation. NGDC Data Annoucement DA 93-MGG-01. Boulder.
J.S., 1994a. Global Ecosystem Framework: Definitions. USGS EROS Data Center Internal Report, Sioux Falls, SD, 37 p.
J.S., 1994b. Global Ecosystem Framework: Translation Strategy. USGS EROS Data Center Internal Report, 39 p.
S., 1999. Pillar of Sand: Can the Irrigation Miracle Last? (New York: W.W.
T., R. Tateishi., Ishiyama, T. and Honda, Y. 1998. Relationships between
percent vegetation cover and vegetation indices, int. j. remote sensing, 1998,
vol. 19, no. 18, 3519- 3535
T., Tateishi, D. 2001. Estimation of percent vegetation cover of grassland in Mongolia using NOAA AVHRR data. GIS development. ACRS 1997, Poster presentation 1.
Rao, C. R.
N. 1993a. Nonlinearity corrections for the thermal infrared channels of the
Advanced Very High Resolution Radiometer: Assessment and recommendations. NOAA
Technical Report NESDIS-69. NOAA/NESDIS. Washington, DC.
Rao, C. R.
N. 1993b. Degradation of the visible and near-infrared channels of the Advanced
Very High Resolution Radiometer on the NOAAP9 spacecraft: Assessment and
recommendations for corrections. NOAA Technical Report NESDIS-70. NOAA/NESDIS. Washington, DC.
S., Agosti, D., Alger, K., Delabie, J. and Musinsky, J. (2001). Examining
fragmentation and loss of primary forest in the Southern Bahian Atlantic Forest
of Brazil with radar imagery. Conservation Biology, 15: 867-875.
S.S. and McDonald, K.C. (1997). Coherent effects in microwave backscattering
models for forest canopies. IEEE Transactions on Geoscience and Remote Sensing,
S.S. and Rignot, E. (1997). Classification of boreal forest cover types using
SAR images. Remote Sensing of Environment, 60: 270-281.
S.S., Nelson, B., Podest, E. and Holt, J. (2000). Mapping land cover types in
the Amazon Basin using 1 km JERS-1. International Journal of Remote Sensing,
S.S., Soares, J.V. and Alves, D.S. (1997). Mapping deforestation and land use
in Amazon rainforest by using SIR-C imagery. Remote Sensing of Environment, 59:
Santa Fe, New Mexico, January 21-26, 1996. National Center for Geographic Information and Analysis, Santa Barbara, California.
Institute (2004). SAS/STAT user?s guide and software release, version 8.2 Ed. Cary, NC: SAS Institute.
J., Staenz K. 2001. Adaptive Threshold for Spectral Matching of Hyperspectral
Data; Canadian Journal of Remote Sensing, Vol. 27, No 3, 2001, pp. 216-224
D., Amarasinghe., U., de Fraiture, C., Keller, A., Molden, D., Sakthivadivel.,
R (eds). 2000. World Water Supply and
Demand: 1995 to 2025, International Water Management Institute, Colombo, Sri Lanka.
J.J. & Drake, N.A. 1993. Linear mixing and the estimation of ground cover
proportions, International Journal of Remote Sensing, 14, 1159-1177
P. 2001. Spectral and hyperspectral analysis with ENVI. ASPRS ENVI user's group
notes. April 22-28, Annual meeting of the American Society of Photogrammetry
and Remote Sensing, St. Louis, Missouri, USA.
Döll, P., Hoogeveen, J., Faures, J., Frenken, K. and Feick, S. 2005. Development
and validation of the global map of irrigation areas. Hydrology and Earth System S
ciences, 9, 535?547
Döll, P., Feick, S., and Hoogeveen, J. 2005, Global map of irrigated areas version 3.0
Johann Wolfgang Goethe University, Frankfurt am Main, Germany / Food and Agriculture
Organization of the United Nations, Rome, Italy.
S., Döll, P., 2001. A digital global irrigated area map ? An update for Latin
America and Europe. Kassel World Water Series 4, Center for Environmental Systems
Research, University of Kassel, Germany, 14 pp + Appendix. [Download pdf, 0.7 MB] . Version 2.1: Update for Africa and Australia,
using the map generation method described in Siebert and Döll (2001)
S., Döll, P., Hoogeveen, J. (2002): Global irrigated area map version 2.1.
Center for Environmental Systems Research, University of Kassel, Germany, and FAO, Rome, Italy. Online documentation and download of 5 min map of irrgated area
S., Döll, P., Hoogeveen, J., 2002. Global
irrigated area map version 2.1. Center for Environmental Systems Research, University of Kassel, Germany, and FAO, Rome, Italy. Online documentation and download of 5 min map of irrgated area
P.M., S.N.V. Kalluri, S.D. Prince, and R.S. DeFries, 1997. The NOAA/NASA
Pathfinder AVHRR 8-km land data set, Photogrammetric Engineering and remote
GangadharaRao, P., Biggs, T., Krishna, M., and Turral, H., 2006. Spectral
Matching Techniques to Determine Historical Land use/Land cover (LULC) and
Irrigated Areas using Time-series AVHRR Pathfinder Datasets in the Krishna
River Basin, India. Photogrammetric Engineering and Remote Sensing, (in press).
P.S., 2005. Inter-sensor relationships between IKONOS and Landsat-7 ETM+ NDVI
data in three ecoregions of Africa. INT. J. REMOTE SENSING, 25 (2): 389-408.
P.S., Enclona, E.A., Ashton, M.S., and Van Der Meer, V. 2005. Accuracy
Assessments of Hyperspectral Waveband Performance for Vegetation Analysis Applications. Remote Sensing of Environment, 91:2-3: 354-376.
P.S., Enclona, E.A., Ashton, M.S., Legg, C., Jean De Dieu, M., 2005. Hyperion,
IKONOS, ALI, and ETM+ sensors in the study of African rainforests. Remote
Sensing of Environment, 90:23-44.
P.S., Gamage, N., and Smakhin, V. 2005. The use of remote sensing data for
drought assessment and monitoring in south west Asia. IWMI Research report #
85. Pp. 25. IWMI, Colombo, Sri Lanka.
P.S., Schull, M., Turral, H. 2005. Ganges and Indus River Basin Land Use/Land
Cover (LULC) and Irrigated Area Mapping using Continuous Streams of MODIS Data.
Remote Sensing of Environment. Remote Sensing of Environment, 95(3): 317-341.
P.S., Stucky, N., Griscom, B.W., Ashton, M.S., Diels, J., Van Der Meer, B.,
and Enclona, E. 2005. Biomass estimations and carbon Stock calculations in the
oil palm plantations of African derived savannas using IKONOS data,
International Journal of Remote Sensing. International Journal of Remote
Tou, J. T.
and R. C. Gonzalez, 1975. Pattern Recognition Principles, Addison-Wesley
Publishing Company, Reading, Massachusetts.
C.J., Grant, D.M., Dykstra, J.D. 2005. NASA?s
Global Orthorectified Landsat Dataset. Photogrammetric Engineering & Remote
Sensing. 70(3), 313-322.
H., 2002. Global irrigated area-working document. Colombo, Sri Lanka: IWMI.
Geological Survey, 1994. Digital
elevation models, data user guide 5. Reston, Virginia, 50 p.
Schilfgaarde, Jan, 1994. Irrigation- a
Blessing or a Curse. Agricultural Water Management 25: 203-219.
K., and S. Jenson, 1996, "Development of continental scale DEMs and
extraction of hydrographic features", proceedings of the Third Conference
on GIS and Environmental Modeling, Santa Fe, New Mexico, NCGIA.
K.L., and Greenlee, S.K., 1996. Development of continental scale digital
elevation models and extraction of hydrographic features. In: Proceedings,
Third International Conference/Workshop on Integrating GIS and Environmental
Xiaoyang Z, Drake, N. and Wainwright, J. 1998.Downscaling
Land Surface Parameters for Global Soil Erosion Estimation Using no Ancillary
Annex 1 : Irrigated Areas of
countries from GIAM10km V 2.0 and other sources
The crop calendars of the classes for Australia differ significantly from rest of the world. Hence the Australian area was
calculated based on the Country crop calendar. However, we have retained the
same irrigated area fractions as global, except for class 19 for which the
fraction was 0.10. Based on this approach, the Country area for Australia is shown in Table A1 below.
Table 1. Irrigated areas of Australia based on GIAM 10-km.
During the next phase, the irrigated areas will be computed
crop calendar; and B.
country-wise crop coefficient.
16.0 Acronyms and
|2 dimensional feature space
other areas of the World segment Advanced
Very High Resolution Radiometer
Board of Irrigation and Power
Active Archive Centers
Terrain Elevation Data
EROS Data Center
greater than 1500 m segment
Resources Digital Analysis System
Resources Observation Systems
Thematic mapper plus
& Agricultural Organization of UN
Forest cover greater than 75 percent
Generic IWMI 628 class map
Land Cover classification for the year 2000
irrigated area map
map of rainfed cropland areas
map of land use/land cover areas
Goddard Space Flight Center
digital elevation model (DEM) with a horizontal grid spacing of 30 arc-seconds
(approximately 1 kilometer)
Geosphere Biosphere Program
Map of the World
Water Management Institute
Water Management Institute Data Storehouse Pathway
clustering algorithm in ERDAS
Earth Resources Satellite-Synthetic Aperture Radar
Photographic Experts Group new imaging compression standard
Resolution Imaging Spectroradiometer
Polar Operational Environmental Satellite System
Aeronautics and Space Administration
Difference Vegetation Index
Environmental Satellite Data and Information System
Geophysical Data Center
Oceanic and Atmospheric Agency
Polar Operational Environmental Satellite System
greater than 2400
Synthetic Aperture Radar
Pour l’Observation de la Terre or Earth-observing Satellites
time spiral curves
Area Available for Irrigation
Observing System (EOS) satellite-NASA flagship satellite under Earth System
less than 280 degree Kelvin
States Geological Survey
and Infrared Imaging Radiometer Suite
Geodetic System 1984
A Global Irrigated Area Mapping
(GIAM) project of this magnitude and complexity can never be done without
substantial and persistent support from many places.
We are very grateful to Prof.
Frank Rijsberman, Director General of IWMI, for his vision, guidance,
intellectual, moral, and financial support. Such support is very rare to come
by. Thank you Frank for your vision and leadership. Dr. David Molden, Principal
Researcher at IWMI, was instrumental in initial funding of the GIAM through
Comprehensive Assessment (CA), A UN Mellennium initiative. We also thank him
for his moral and intellectual inputs. The project was initially conceptualized
and lead throughout by Hugh Turral. Hugh?s energy has been amazing given that
he has to juggle this with Theme leadership at IWMI. But he has always had time
for sustained and stimulating discussions on GIAM. That is true leadership. Sarath
Abayawardena, former Global Research Director (GRD), was instrumental in laying
a strong foundation for the RS\GIS laboratory at IWMI. We remember that support
very much. Julie Van der Bliek, the present GRD head, has continued this
support and has steered us towards a spatial data policy at IWMI. All of these
efforts have helped GIAM. Honestly, we can not list everyone who have helped us
at various stages of the project in many different ways. Everyone in RS\GIS
unit have helped in one way or the other. Thank you Aminul Islam for the
excellent ground truth data of the world and for the rainfall data compilation.
Wasantha Kulawardana for support when needed. Sarath Gunasinge, and Ranjith
Alankara are always there for silent strong support in producing maps and flow
charts. Jacintha Navaratne provided outstanding secretarial services, most
IWMI India office was specially
helpful. Thank you Trent Biggs, Muralikrishna, and Parthasarathi. We would like
to thank the Food and Agricultural Organization (FAO)\University of Frankfurt
(UF) for lively discussions on the two Global Irrigated Area Maps (FAO\UF and
IWMI). Specifically, we would like to thank Stefan Siebert, Jippe Hoogeveen,
The NASA Goddard Space Flight
Center (GSFC) made available the AVHRR time series used in this work. Special thanks
to Dr. Ron Smith and group. The Landsat data were downloaded from the University of Maryland?s Global Land Cover Facility (GLCF). Several datasets such as the
GTOPO30 1-km and SRTM 90 meter elevation data are downloaded from the USGS\EROS.
The forest cover data from Dr. Ruth DeFries of University of Maryland. Rainfall data was provided by Dr. Tim Mitchell of East Anglica Climate Research
Group. The JERS SAR data from Saatchi and group. The volunteer ground truth
data from degree confluence project was invaluable. The Google Earth Data is
state-of-art and was widely used. Without these great datasets, made available
for free, the project would never have got started. So we are very, very
grateful to these Agencies and numerous people behind it.
At times like this, the release
of the first satellite sensor based Global Irrigated area map, we cherish the
wonderful memories of our revered Gurus and venerable universities; from whom
we learnt and got taste of knowledge. They were true giants to whom this credit
should really go. It is unfair to mention one name and not several others. But
it will at least take a page to do justice to everyone.
Finally, the entire team (see
list of authors and also in acknowledgements) that worked on this project has
been wonderful. Long hours and stiff deadlines were common. At times, our
patience and resolve were tested severely due to nightmares of data chaos and
organization. But the morale and motivation was always high. Support for each
other as good as it can get. Intellectually very stimulating and challenging-
something we always enjoyed and looked forward to. In many ways, we all learnt
and made progress.
 International Water Management Institute (IWMI), Colombo, Sri Lanka; 2Boston University, USA;
List of Figures
||Processing chain for the
Global irrigated area map (GIAM)
used in GIAM. The mega-file of 159 layers of data and consists of 144 AVHRR
10-km monthly layers from 3 years, 12 SPOT monthly layers from1999 year, single
layer of DEM, mean annual rainfall from 40-years, and forest cover
||Primary and secondary data sets used in the Mega-file.
||JERS-1 SAR 100-m image tile mosaicks for the Central Africa. The rainforests of the Africa and the Central America were studied using JERS-1 SAR 100-m data for two periods in
ETM+ 150-m images of the World as ?ground-truth?. The Landsat ETM+ (Geocover
2000) orthorectified images for the nominal year 2000 at 150-m resolution were
used as a ?ground-truth?
(GT) data of the World by IWMI. Groundtruth data assembled from multiple
locations and times by IWMI projects and staff.
data of the World from the Degree Confluence Project (DCP)
||Summary of analysis to determine irrigation land use classes (Part 1).
||Summary of analysis to determine
irrigation land use classes (Part 2).
less than 360 mm segment (PLT360-segment). These arid or semi-arid areas
provide distinct contrast between areas with and without vegetation
||Forest density greater than 75 percent
(FGT75-segment). These areas have low probability of agriculture, except in
rare fragments of slash and burn
||Time-series AVHRR 10-km profile of spectral classes is illustrated for AOAW-segment. The AOAW-segment initially had 350 classes. The plot of some of these classes highlights the spectral characteristics of each class. A quantitative approach to determine which of these classes match is performed through SCS R2-squared (e.g., Table 4)
||Identifying similar irrigated classes using spectral matching. Spectral matching in combination with ground truthing and ideal spectra helped group similar irrigated double crops (shown in red, for classes 50, 59, 60, 67, and 74). The same logic was used to group: wetland crops (sown in blue; class numbers 10 to 15), and continuous irrigation (shown in green; classes 2, 6, 7, 23, and 24)
||The process of combining classes in spectral matching techniques (SMTs) is illustrated. First, the SCS R2-values are determined for a matrix of classes. The time-series spectra of classes with high SCS R2-values are then matched. Grouped classes are investigated further using all other types of information including groundtruth. This leads to distinct groups such as: boreal forests and tropical forests. Finally, the classes of similar types are color coded
||The process of spectral matching
techniques (SMTs) is illustrated. The 17 classes considered in Figure 17 are
further refined by quantitative and qualitative SMTs that lead to 3 distinct
||The spectral similarity value
(SSV) to match spectra. In this figure, unsupervised class spectra are compared
with ideal spectra of distinct irrigated classes: (a) major irrigation in
Ganges basin (ideal spectra in red, unsupervised grouped class spectra in
magenta), (b) supplemental irrigation from mid-west USA (pivot sprinkler) and
Syria (underground water) (ideal spectra in light blue and actual unsupervised
class spectra in deep blue), and (c) delta irrigation from Bangladesh (ideal
spectra in light green, unsupervised class spectra in deep green). The smaller
the SSV, greater the match in shape and magnitude
||Figure 20a. Google Earth ?zoom
in? views to identify a class. One preliminary class is spread out across the
world. The class was investigated using 50 Google sample points that were
randomly chosen. The figure shows the spread of the class across the world and
Google Earth hi-res image at 2 locations: center pivot ground water irrigation
in the USA and surface irrigation in Sudan
||Brightness-greenness-wetness (BGW) plot
||Space-time spiral curve (ST-SCs)
to track class changes in 2-dimensional(2-d) space and time
NDVI spectral profile to identify and delineate classes
||AVHRR derived skin temperature
versus AVHRR NDVI for semi-arid and tropical crops
||Decision trees to resolve
mixed classes. Forest cover density (%) is used to resolve mixed class # 13 in
the precipitation > 2400mm per year segment
||Decision tree rules to resolve mixed classes. The Maximum, minimum, and average NDVI were used, in a decision tree framework, to separate out distinct areas within class 17.
||Geocover Landsat 150-m data of the World in class identification and labeling process
||Irrigated areas and other LULC in the Ganges basin India. Irrigation in the Ganges includes tube wells in alluvial areas,
reservoirs, and river diversions.
||Irrigated areas and LULC classes from different parts of the World
||Irrigated areas and LULC classes from different parts of the World from the degree confluence project
||Class naming convention. The standardized class naming convention is depicted in this figure. At different levels, the class naming may or may not include a particular category such as scale of irrigation or the intensity.
||Summary of area abstraction from the 28 irrigation class map
||Irrigated area by Google earth estimate (GEE). For each GIAM10km-28 classes Google earth estimates (GEE) of irrigated area fraction (IAF) were estimated using Google earth images. Thirty points were taken for each class and averaged. The fraction calculation for one class is illustrated.
||Irrigated area fraction from high resolution imagery (IAF-HRI). For each of the GIAM10km-28 Classes, The IAF-HRI were estimated by masking Landsat images for the area occupied by the class and then determining irrigated vs. non-irrigated areas
||Sub-pixel de-composition technique (SP-DCT).
||Relationship between percent irrigated area of class 1-20 and the AVHRR
NDVI computed using band 1max and AVHRR band 2max reflectivity
||GIAM10km V2.0 28 class Map
||GIAM10km V2.0 8 class Map
||GIAM10km V2.0 3 class Map
||Trends in irrigated area since 1800. The IWMI
at the end of the last millennium not only considered area irrigated, but also
the intensity (i.e., area irrigated during different seasons in 12-month
period) and informal irrigation (e.g., ground water, tanks). This gives an
estimate of 263 million hectares during ?main? cropping season (Season 1) and a
total of 480 million hectares from 3 seasons: First Crop (263 Mha), Second Crop
(176 Mha), and Continuous crop (41 Mha).
global irrigated area class snap-shot illustrations for GIAM classes. The
snap-shots (e.g., photos, high-resolution images) of 4 distinct classes for a
GIAM10km V2.0 class
||Figure 38. Evaluation of the GIAM for large scale, small scale, informal, and supplemental irrigation. The IWMI GIAM and India?s Central board of irrigation and power (CBIP) irrigated area maps are evaluated for: (a) large scale irrigation- (Figure 38a and 38b); (b) informal irrigation such as ground water and tanks (Figure 38c and 38d); and (c) small scale (e.g., minor reservoirs) irrigation (Figure 38e and 38f)
||Comparison of the two global irrigated area maps: GIAM10km V2.0 and FAO/UF V3.0
||Comparison of irrigated areas of 40 leading countries between IWMI GIAM10km V2.0 vs. FAO\UF V3.0
||Single crop (red) and double crop (cyan) irrigation in lower Ganges
||Double crop (left) and single crop (right) irrigation in Zahandeh and Rud
||Evaluation of GIAM for conjunctive irrigation. The rainfed class with significant central pivot supplemental irrigation in the Pampas in Argentina
List of Tables
||Characteristics of the Satellite sensor and
secondary datasets used in mapping Global irrigated areas. These datasets were
Compiled into a 159-band layer stack
used in conjunction with the megafile.
locations of the ideal target spectra for 7 irrigated area classes
||The SCS R2-value
matrix of spectral classes
||Sample characteristics of IWMI-951 class generic
||Indicative class name through use of secondary data
||Process of aggregation of classes from the generic map. The irrigated area classes were aggregated from 951 class map based on the methods discussed in sections 5, 6, and 7. Similar approach was used to aggregated classes into 28 or 8 or 3 class map
||Characteristics of irrigated areas. Intensity and cropping calendar for the GIAM classes in India
||Irrigated areas of the World from the GIAM10km-28 classes V2.0 map using IAF from HRI and SPDT. The irrigated areas of the world are calculated from the GIAM10km V2.0 map based on the cropping intensity. The class-wise irrigated area details are shown for GIAM10km-28 classes
||Irrigated areas of the World from the GIAM10km-8 classes V2.0 map using IAF from HRI and SPDT. The irrigated areas of the world are calculated from the GIAM10km V2.0 map based on the cropping intensity. The class-wise irrigated area details are shown for GIAM10km- 28 classes
||Irrigated areas of the World from the GIAM10km-3 classes V2.0 map using IAF from HRI and SPDT. The irrigated areas of the world are calculated from the GIAM10mn V2.0 map based on the cropping intensity. The class-wise irrigated area details are shown for GIAM10km- 3 classes
||Irrigated areas of the Continents. The GIAM10km continental areas are compared with the FAO Aquastat and the National statistics
||Irrigated areas of the Countries. The GIAM10km country areas are compared with the FAO Aquastat and the National statistics
||Irrigated areas of the river basins. The GIAM10km river basin areas are compared with the FAO Aquastat and the National statistics
||Accuracy assessment of IWMI GIAM V2.0 Vs. FAO/UF V3.0 vs. CBIP using ground truth data. The IWMI Global irrigated area map (GIAM) is compared with the: (a) Global irrigated area map of the FAO/Frankfurt University, and (b) India's Central Board of Irrigation and Power (CBIP)
||Accuracy assessment of IWMI GIAM V2.0 Vs. FAO/UF V3.0 vs. CBIP using google earth ground truth (GEGT). The IWMI Global irrigated area map (GIAM) is compared with the: (a) Global irrigated area map of the FAO/Frankfurt University, and (b) India's Central Board of Irrigation and Power (CBIP).