Estimating the potential of rainfed agriculture in India: Prospects for water productivity improvements

This article appeared in a journal published by Elsevier. The attachedcopy is furnished to the author for internal non-commercial researchand education use, including for instruction at the authors institution

and sharing with colleagues.

Other uses, including reproduction and distribution, or selling orlicensing copies, or posting to personal, institutional or third party

websites are prohibited.

In most cases authors are permitted to post their version of thearticle (e.g. in Word or Tex form) to their personal website orinstitutional repository. Authors requiring further information

regarding Elsevier’s archiving and manuscript policies areencouraged to visit:

http://www.elsevier.com/copyright

Author's personal copy

Integrating remote sensing, census and weather data for an assessment of riceyield, water consumption and water productivity in the Indo-Gangetic river basin

X.L. Cai a,*, B.R. Sharma b

a International Water Management Institute, 127 Sunil Mawatha, Pelawatta, Battramulla, Sri Lankab International Water Management Institute, Delhi Office, NASC Complex, CG Block, Dev Prakash Shastri Marg, Pusa, New Delhi 110012, India

1. Introduction

Food security is a major concern in many parts of the world whichis also the case in countries within the Indo-Gangetic Basin (IGB),where significant expansion of crop cultivation is not feasible. Tomeet the rising food demand by both the increasing population andthe changing diet patterns, the world needs to ensure sustainableland productivity improvement over the coming decades (Moldenet al., 2007). Among the many constraining factors of landproductivity such as soil, seed, fertilizer, insects and diseases, wateris a key constraint. With the ever-competitive demand fromindustry, domestic users and the ecosystem, the agriculture sectoris seen to get reduced water allocation despite the increasingpressure for more food production (Rosegrant et al., 1997). Together,the increasing food demand and decreasing water allocation suggestthat the agriculture sector has to produce more food with less water,that is, to increase the WP of agriculture.

Water is one of the most critical inputs to agriculture.However, the level of water use differs significantly across

regions, farming systems, canal command areas, and even farmplots (Molden et al., 2003). The differences come from manyaspects. In irrigated areas, these include storage management,the source of irrigation water, the timing and efficiency ofirrigation water delivery, on-farm management practices andreuse of return flow. In rain-fed areas, these include the croppingpattern, resources conservation technologies, rainwater harvest-ing, and integrated watershed management. Each of theseaspects involves a number of interventions to meet crop waterdemand. However, it is not clear how water is better used bycrops and contributes to improved productivity with all theseefforts, especially in large river basins across countries. Regionalor basin-wide water accounting and WP analyses are required tomeasure the effectiveness of these interventions and to under-stand system water balance components and water input–agricultural output relation.

Remote sensing (RS) is an innovative tool to observe landsurface processes on a large scale and in a cost-effective approach(Schmugge et al., 2002). Bastiaanssen et al. (1999) firstly usedremotely sensed data to estimate both crop yield and evapo-transpiration for a WP study in the Bhakra command area, India.This approach was further developed by using the SEBAL methodalone (Zwart and Bastiaanssen, 2007). Numerous studies haveseparately demonstrated the strength of RS in estimating crop

Agricultural Water Management 97 (2010) 309–316

A R T I C L E I N F O

Article history:

Received 21 April 2009

Accepted 26 September 2009

Available online 28 October 2009

Keywords:

Census data

ET

Indo-Gangetic Basin

Remote sensing

Rice

Water productivity

A B S T R A C T

Crop consumptive water use and productivity are key elements to understand basin water management

performance. This article presents a simplified approach to map rice (Oryza sativa L.) water consumption,

yield, and water productivity (WP) in the Indo-Gangetic Basin (IGB) by combining remotely sensed

imagery, national census and meteorological data. The statistical rice cropped area and production data

were synthesized to calculate district-level land productivity, which is then further extrapolated to

pixel-level values using MODIS NDVI product based on a crop dominance map. The water consumption

by actual evapotranspiration is estimated with Simplified Surface Energy Balance (SSEB) model taking

meteorological data and MODIS land surface temperature products as inputs. WP maps are then

generated by dividing the rice productivity map with the seasonal actual evapotranspiration (ET) map.

The average rice yields for Pakistan, India, Nepal and Bangladesh in the basin are 2.60, 2.53, 3.54 and

2.75 tons/ha, respectively. The average rice ET is 416 mm, accounting for only 68.2% of potential ET. The

average WP of rice is 0.74 kg/m3. The WP generally varies with the trends of yield variation. A

comparative analysis of ET, yield, rainfall and WP maps indicates greater scope for improvement of the

downstream areas of the Ganges basin. The method proposed is simple, with satisfactory accuracy, and

can be easily applied elsewhere.

� 2009 Elsevier B.V. All rights reserved.

* Corresponding author at: International Water Management Institute, PO Box

2075, Colombo, Sri Lanka. Tel.: +94 11 2880000; fax: +94 11 2786854.

E-mail address: [email protected] (X.L. Cai).

Contents lists available at ScienceDirect

Agricultural Water Management

journa l homepage: www.e lsev ier .com/ locate /agwat

0378-3774/$ – see front matter � 2009 Elsevier B.V. All rights reserved.

doi:10.1016/j.agwat.2009.09.021

Author's personal copy

yield (Yang et al., 2008) and in monitoring crop consumptive useof water (evapotranspiration) (Courault et al., 2005), which arekey elements of crop water consumption and WP studies. Theseapproaches require relatively less ground information whileproviding vital inputs to enable a more comprehensive analysis.Courault et al. (2005) categorized various approaches of ET mappinginto four groups: (a) empirical direct methods, (b) residual methodsof the energy budget, (c) deterministic methods, and (d) vegetationindex methods. The several operational models commonly used,such as Surface Energy Balance Algorithm for Land (SEBAL)(Bastiaanssen et al., 1998), Mapping Evapotranspiration at HighResolution using Internalized Calibration (METRIC) (Allen et al.,2007) and Simplified Surface Energy Balance Index (S-SEBI) (Roerinket al., 2000), are based on the residual component calculationcombining empirical and physical modules to facilitate simpleparameterization on a large scale where data constraints are a majorconcern. Applications showed promising accuracy with limitedground information requirements. Vegetation index methods areadapted from the FAO reference ET and crop coefficient (Kc)approach. Multispectral vegetation indices are used to derive thecrop coefficient while reference ET is calculated from meteorologicaldata (e.g., Gonzalez and Mateos, 2008). These methods avoid thestill complex parameter estimating processes in energy balancemodels which require enriched processing experiences. However,the relation between spectral vegetation indices and crop coeffi-cients in different locations are uncertain and thus always subject tomodification and validation prior to new applications. Moreover,these methods require the use of crop maps to determine ET, whileenergy balance models estimate ET independent of land use.

Crop biomass and yield can be assessed from space usingvegetation indices. For example, the accumulative photosynthe-tically active radiation absorbed by plants (APAR) could beexamined by satellite sensor, which is then linked to biomassaccumulation and, consequently, yield (Bastiaanssen and Ali,2003). Among the numerous remote sensing crop yield monitoringand predicating methods, linear equations, which link theNormalized Difference Vegetation Index (NDVI) to crop yield,are proven to be simple, yet successful methods (Groten, 1993). Forrice, Thiruvengadachari and Sakthivadivel (1997) found thehighest correlation at the heading stage with the coefficient ofdetermination statistically significant at 0.76 in India. Such studiestry to build an empirical quantitative relation between grain yieldand vegetation index that, however, is highly variable and thusonly applicable to the specific area (Moulin et al., 1998).

Another issue in the application of RS is the often-seen gapbetween RS interpretation and national census which furthercomplicates the results, and effectively prevents stakeholders frommaking better use of RS techniques (Frolking et al., 2002). In manycases, statistical data are used to validate RS results at the finalstage (e.g., Frolking et al., 1999; McCoy, 2004). Some researcherstry to combine census data and remote sensing imagery in theinterpretation processes. These efforts are frequently seen inpopulation estimation (e.g., Sutton et al., 2001; Harvey, 2002), landuse or land cover mapping (Mesev, 1998; Hurtt et al., 2001;Frolking et al., 2002; Neto and Hamburger, 2009) and the drivers ofchange (Rindfuss and Stern, 1998). However, crop productivityestimates are often achieved through remote sensing interpreta-tion and then validated with ground measurements.

This article presents an innovative approach to combinemeteorological data, ground survey, and national census withremotely sensed imagery to assess rice water consumption (actualET), rice productivity, and finally crop WP for the large IGB in SouthAsia. The statistical data were synthesized to calculate district-/state-level land productivity, which is then further extrapolated topixel-level values using a MODIS NDVI image, based on a cropdominance map. The actual ET is estimated with an SSEB model

taking meteorological data and MODIS land surface temperature(LST) products as inputs. WP maps are then generated by dividingthe crop productivity maps by ET maps.

2. Study area

The IGB, also known separately as Indus and Ganges riverbasins, covers a huge area of 2.25 million km2 in Nepal, significantparts of India, Pakistan, Bangladesh and small parts of China andAfghanistan (Fig. 1). Diverse climatic, topographic and soilconditions exist in the basin. The climate is strongly characterizedby monsoons with annual average precipitation varying from lessthan 100 mm to 4000 mm, most of which occurs from June toOctober. The basin could be classified into three physiographicregions: mountain areas, plains and deltas. The mountain areasoriginate from the southern slope of the Himalayas and graduallyextend to two deltas: towards the southwest till the Arabian Seaand towards the southeast till the Bay of Bengal. IGB is the world’smost populous basin with a population of 747 million (2001), ofwhich around three-quarters are living in rural areas. The fourmajor countries, India, Pakistan, Bangladesh and Nepal, are allexperiencing fast population growth which imposes high pressureon water and food security.

Out of the total drainage area, more than 50% of the entire IGB iscultivated. Rice–wheat rotation is the predominant croppingsystem in the region, mixed with cotton, sugarcane, pulses, milletand jute, etc. Extensive irrigation is practiced in the major foodproduction zones for kharif paddy rice and rabi wheat along withother crops. A large quantity of surface water is diverted forirrigation through canals. However, groundwater use throughnumerous pumps is the most popular practice in India andBangladesh (Shah et al., 2006), where it is difficult to monitor andcount the water uses and depletion. Crop-sowing dates and growthperiod vary according to climatic conditions, water availability,farmers’ agronomic practices and crop varieties (Ullah et al., 2001).The selection of appropriate transplanting and harvest dates of riceis important to assess crop water use. The ET around the riceharvest period is usually very low because of the low canopyphotosynthetic rate, dried soils and dry weather conditions. Hence,several days more or less would not affect much the summed ET,and subsequently the WP. The selection of the transplanting date ismore critical as there is standing water before and aftertransplanting; and pan evaporation reaches the maximum valuein June at the rate of around 8 mm per day (Mahajan et al., 2009).Oza et al. (2008) used microwave remote sensing data to monitorthe rice growth stage in India, and found that most of the rice wastransplanted around June 11. The study period for analyzing rice

Fig. 1. The Indo-Gangetic river basin study area.

X.L. Cai, B.R. Sharma / Agricultural Water Management 97 (2010) 309–316310

Author's personal copy

yield and ET was then set to June 10 through October 15 (128 days,corresponding to 16 MODIS 8-day LST images).

3. Methodology

3.1. Crop dominance map

Crop-specific water consumption and productivity analysisrequire a knowledge of the physical location of the crop. Threeland use maps were collected whose characteristics are listed inTable 1. These maps were produced from various sensors andhave different purposes and interpretation techniques. A quickinvestigation indicated significant differences between the maps.Spatial analysis and class identification techniques as proposed byThenkabail et al. (2007) were used to synthesize the three maps toan improved rice–wheat dominance map taking census data andgroundtruth information as input. A detailed description isbeyond the scope of this article but it could be found in Cai andSharma (2009).

The groundtruth mission was carried out in India in October2008. The mission collected 175 samples scattered in the states ofHaryana, Rajasthan, Punjab, Himachal Pradesh and Uttar Pradesh.Detailed land use patterns were recorded including a crop mixturepercentage eye estimate, crop growth period and past crop types.Data on crop yield, watering sources and irrigation intensity werecollected from farmers through a questionnaire survey. Cut andweighing methods were also applied in selected rice fields tocollect data on above-ground biomass and grain yield. Theinformation on yield was used in rice productivity mappingdescribed in the next section.

3.2. Rice yield mapping

A rice yield map was produced in two steps. First, a districtaverage yield map was generated from census data with vectors ofadministrative boundaries. Then the district yield map wasextrapolated to pixel level by using MODIS NDVI products of therice heading stage.

The district yield map is a straightforward way to produce as cropproduction data are often collected and/or reported at the admin-istrative boundary level. However, the actual water consumption andcrop performance are dependent on topography, soil, water, climateand on-farm management practices, which do not necessarilycorrespond to administrative boundaries. Hence, there is a need toidentify the actual boundaries of crop performance variations, whichcould be easily assessed from pixel-based raster maps. In this study,the district-level rice productivity map was further disaggregated tothe pixel level using MODIS NDVI data as a bridge.

The national census data were collected from country sources.The data set includes district-level crop area, yield and productionfor India, Pakistan, Bangladesh and Nepal. The time window of thedata mainly focused on year 2005–2006 with some data in a fewdistricts varying from 2001 to 2004. Yield and production of thesedistricts were then estimated for 2005–2006 using state-level datausing linear algorithm assuming a fixed percentage of districtproduction to state-level production.

The main period of paddy rice growth in the IGB is from middleof June to early October. Assuming an average of 80 days aftertransplanting (Oza et al., 2008), the rice heading stage isdetermined to be from 29th August to 5th September. EightMODIS NDVI images, one day each, were downloaded from NASAWarehouse Inventory Search Tool (WIST) (http://wist.echo.nasa.gov/api/). To ease the heavy clouds around this period, the imageswere compared pixel by pixel to extract maximum values into anMVC (Maximum Value Composition) image which eliminated thepixels with low NDVI values as a result of any cloud effect. Non-rice-dominant areas are masked out using the crop dominancemap. District average NDVI values for rice are then calculated andrelated to district average rice yield. In this way the linearregression equation was built up as shown in Eq. (1):

Yieldp ¼ Yieldavg �NDVIP

NDVIavg(1)

where Yieldp and Yieldavg are average yields of an individual pixeland district, respectively, NDVIavg is district average NDVI andNDVIP is NDVI of any given pixel during the heading stage. Theequation was then applied to each pixel on the NDVI MVC ricesubset, leading to a 250 m � 250 m resolution yield map of rice(kharif).

3.3. ET estimation

The SSEB model (Senay et al., 2007) combines remote sensingand meteorological data to estimate actual ET. Based on the sameassumptions in SEBAL and METRIC models, SSEB assumes thetemperature differences between land surface and near-surfaceair, which vary linearly with LST. However, this assumption isfurther extended by stating that latent heat flux (actualevapotranspiration) also varies linearly with LST. Hot pixels andcold pixels were used to represent ‘‘no ET’’ or ‘‘maximum ET.’’Therefore, the actual ET (ETa) of other pixels is linearly distributedbetween the range of hot pixel (ETa = 0) and cold pixel(ETa = maximum ET), resulting in a proportional ET fraction value(ETf) for each pixel as expressed in Eq. (2):

ETf ¼TH � TX

TH � TC(2)

where ETf ranges from 0 to 1, TH and TC are the temperatures of hotand cold pixels, respectively, and TX is the surface temperature ofany pixel on the image. Using this equation we can generate an ETfraction map based on the LST map from thermal imagery.However, apart from the hot pixels with no ET, another ‘‘anchorpixel’’ needs to be identified to determine the slope. In the SSEB it isproposed to adopt reference ET (ET0) calculated from weather dataas maximum ET corresponding to cold pixels. The ETa of day i (ETa,i)can be generated by multiplying ETf with ET0 of day i as shown inEq. (3):

ETa;i ¼ ET0 � ETf (3)

where ET0 is the actual ET for reference vegetation (alfalfa or grass)under certain conditions. The ETa of various crops could be more orless than ET0 after multiplying it with crop coefficient (Kc) and soil

Table 1Characteristics of existing land use and land cover maps used to produce crop dominance map.

Name Nominal

year

Data

source

Resolution

(m)

Number of

classesa

Agricultural

classesa

Reference

Global land cover characteristics data base (GLCCD) 1992–1993 AVHRR 1000 24 5 Loveland et al. (2000)

Global irrigated area mapping (GIAM) 2003 MODIS 500 30 30 Thenkabail et al. (2006)

Paddy rice map 2002 MODIS 500 1 1 Xiao et al. (2006)

a Note: The number of classes fall into IGB study area.

X.L. Cai, B.R. Sharma / Agricultural Water Management 97 (2010) 309–316 311

Author's personal copy

water stress factor (Ks). Hence, this study modified the SSEB modelby using ETp which is defined as ET0 multiplied by Kc instead of ET0.ETp is the maximum crop evapotranspiration under ideal crop-growing conditions that, we assume, exist in some spots of asystem that is not too poor.

Ideally, ETp should be as explicit in the spatial domain as ETf.However, this is not practical as weather data are collected fromindividual stations. While data from 54 weather stations scatteredthroughout the basin were collected, these data had items only onmaximum/minimum/mean temperature, mean humidity andwind speed. Hence, Hargreaves’ equation was used to calculatedaily ET0. Kc values of different rice growth stages were extractedfrom Ullah et al. (2001). Daily ETp raster maps were extrapolatedfrom point ETp using the tension spline algorithm.

Each ETf image represents an 8-day period. The underlyingassumption is that while daily ET0 varies, ETf remains constantthroughout the 8 days, which is acceptable when no significantland cover changes, such as crop harvest. Daily ETa,i was thencalculated and summed up to generate seasonal ET (ETa,s) maps.

MODIS 8-day LST products were downloaded through NASAWIST. As the data set has gone through several calibrationprocesses it is directly used in this study. The original images weremerged into basin mosaics and converted to ERDAS imagine formatwith the WGS 84 geographic projection system. During thisprocess, the digital numbers were also converted into LST valuesusing the scale factor provided in the user’s manuals. Individualmosaics of time series images were then further layer-stacked intoone image file ready for use. Cloudy areas in kharif, 2005 wereeliminated by replacing the values with the average of two images,one before and one after, in the nearest clear dates. Over large scale,LST declines with the increment of altitude. An average lapse rateof 6 8C per 1000 m by Thayyen et al. (2005) in the same region is

used to correct LST by overlapping the LST images with basindigital elevation model (DEM). Daily weather data from 54 stationswere collected for years 1995–2007. Data gaps in 2005–2006 werefixed using average values of the same Julian date as in other years.

3.4. WP mapping

Units differences in yield map (tons/ha) and ET map (mm) wereadjusted to kg/m2 and m3/m2, respectively, to calculate the WPindicator (kg/m3). The resolution of ET and crop dominance mapswas rescaled from 1000 m and 500 m to 250 m to match the cropyield map. The WP map was then produced by dividing the riceproductivity map by the seasonal ET map for the paddy rice area.

4. Results

4.1. Rice yield map

The rice productivity map is shown in Fig. 2. The average riceyield for Pakistan, India, Nepal and Bangladesh parts of the IGB is2.60, 2.53, 3.54 and 2.75 tons/ha, respectively. However, tremen-dous differences exist in different areas of the basin. The IndianPunjab state with some adjacent areas from Haryana and Rajasthanstates (red patch in Fig. 2 rice yield map) has an average yield of6.18 tons/ha, which is significantly higher than that in most otherareas within the basin. (For interpretation of the references to colorin this sentence, the reader is referred to the web version of thearticle.) The low-yield rice areas are also mainly found in the Indianstates of Madhya Pradesh, Rajasthan and Bihar with average yieldsof 1.18, 1.49 and 2.04 tons/ha, respectively. This explains thereason for the low average yield of India. With the spatially explicitmap of rice yield presented for each pixel, significant variability is

Fig. 2. Administrative boundary, actual ET (ETa), yield and water productivity of rice in IGB for year 2005.

X.L. Cai, B.R. Sharma / Agricultural Water Management 97 (2010) 309–316312

Author's personal copy

also observed at the local scale. For example, the average yield ofIndian Punjab is 6.18 tons/ha; yet, it has around 1% area with arelatively low yield of less than 3 tons/ha. Although the averageperformance in Bihar, India is very poor, it has a relatively high-yield area (4 tons/ha) in a 37-km-diameter circle centered at25.4N, 84.44E (southwest of Bhojpur district).

4.2. ET map

The rice ETa map and the histogram distribution are shown inFigs. 2 and 3, respectively. The seasonal average paddy rice ETa

from June 10th to October 15th, 2005 is 416 mm, ranging from167 mm to 608 mm with a standard deviation of 104.6 mm (1%points were sieved). The average value is significantly less than thereference ET (558 mm) and the rice potential ET (610 mm). Theaverage ET for non-rice croplands of the same period is 345 mm,with a slightly higher standard deviation of 109.4 mm.

Significant variation of ET is observed in Figs. 2 and 3. Theadjoining area of the Indus and Ganges catchments in Punjab,Haryana and west Uttar Pradesh, India, covering 7.9% of the totalrice area, has the highest evapotranspiration with an average valueof 551 mm. The northern part of West Bengal also has a high ET of528 mm. A high-ET belt occurs from the Khulna division ofBangladesh to the Indian states of West Bengal, northern Bihar,central Uttar Pradesh, Haryana and Punjab (see Fig. 2). Low ETareas are mainly found in the Indian states of Madhya Pradesh andRajasthan, which are far from the main stream of the Ganges river,and the southern part of Punjab and the northern part of Sindprovinces in Pakistan, where more mixed cropping pattern isobserved. Overall, the ET variations displayed similar trends asshown by yield, although in parts of Bihar state the relatively highET is accompanied by low yield.

4.3. WP map

The average rice WP for the basin is 0.74 kg/m3, with minimum,maximum and standard deviation values of 0.18, 1.8, and 0.329 kg/m3, respectively, (1% extreme pixels sieved). The WP variationgenerally closely follows the pattern of yield variation. The IndianPunjab and adjoining areas, covering 6% of the total rice area, havea very high WP with an average value of 1.32 kg/m3. However, asmuch as 23% of the total rice areas have a WP less than 0.5 kg/m3

occurring mainly in Madhya Pradesh and Bihar states of India andthe Dhaka division of Bangladesh.

Some areas show different trends in WP variation compared tothe yield and ET maps. A high WP strip, around 10–70 km in width,originates from 75.5E, 29N in the southern Haryana state and goestowards the east till the southern Bihar state, India (85.2E, 24N).

The yield for this area is relatively low with an average value of3.2 tons/ha. However, the average ET of the same area is as low as316 mm, making the WP relatively higher. Possibly, this stretchhas the largest area under rain-fed rice, where monsoon rains areable to only partially meet the rice ET needs. The higher WP valueshere do not suggest a satisfactory performance in this case. Rather,it provides interesting clues to reveal the reasons for thedifferences, the potential for yield improvement or ET reduction,and the possible interventions by ‘‘scaling up’’ to other areas.

4.4. Accuracy analysis

Due to data limitation for the large basin area, in this section wediscuss the accuracy of yield, ET and WP in both quantitative andqualitative approaches.

The estimated yields generally agree with the actual yieldsmeasured from the field campaign with a coefficient ofdetermination (R2) value at 0.44 (Fig. 4). One of the constraintsto be considered in this comparison is that the MODIS NDVI spatialresolution is 250 m � 250 m, while the measured yield is takenonly from a 1 m � 1 m plot. Although the sample was taken torepresent rice conditions of the nearby area, it remains uncertainto what degree the measured yield is close to the average yield ofthe MODIS pixel area. However, the overall accuracy of the finalyield map is bounded in a satisfactory range because of tworeasons. First, the model is a linear extrapolation of district data.Hence, it depends mostly on the quality of input data taken fromthe census which is governed by the guidance of officiallyaccepted national standards. Second, possible statistical errors ofcertain districts cause errors only within the district boundariesand would not be transferred to other areas. Table 2 listed theyield, ET and WP values generated from this study and extractedfrom literature for the IGB. The yields generally fall in the rangegiven by literature. But the values are significantly lower in theMadhya Pradesh state, India. This is because the yields providedby Mishra et al. (1997) were recorded from controlled fieldexperiments, while statistical values show generally low valuesbut high variation within the state.

The ET map is more difficult to validate due to the highvariability of actual ET and the low resolution of the ET mapproduced from MODIS LST 1 km products. The SSEB model hasbeen applied and validated in Bangladesh (Senay et al., 2007),China (Cai and Cui, 2009) and central Asia (Cai et al., 2009). Itshowed a similar accuracy with SEBAL algorithm except thathigher values for water bodies were found in the SSEB model. Itwas also found that the model works particularly well in thevegetative area including rice agriculture, which is the majorinterest of our studies. Table 2 compares the ranges of ET valuesfrom this study and from literature. The ET from this study isslightly lower in most cases. This is because the spatial resolution

Fig. 3. The histogram distribution of rice and non-rice cropland ET in IGB for the

period of 10 June–15 October, 2005. Fig. 4. Relation between field-measured and modeled yield of rice.

X.L. Cai, B.R. Sharma / Agricultural Water Management 97 (2010) 309–316 313

Author's personal copy

of the ET map is 1 km � 1 km. In such a big area, rice only grows in apart of it. The rest could be field bunds, trees, barren land andother crops. Depending on the percentage of the rice area, theaverage ET of the pixel could be as low as 167 mm. A similartrend was also observed from past studies. Under the sameproject framework, Bastiaanssen et al. (2006) and Singh et al.(2006) examined WP of rice in the Sirsa district, India, usingremote sensing and SWAP modeling separately. The ET, yieldand WP level by remote sensing are all lower than point SWAPmodeling. And this study which used 1 km MODIS data also haslower ET and WP values than those given by Bastiaanssen et al.which are based on the fusion of the ETM+ (30 m) and theAdvanced Very High Resolution Radiometer (AVHRR, 1.1 km).The groudtruth mission has made eye estimates on thepercentage of rice areas within each 60 m � 60 m square. Thedata showed a higher percentage of rice areas (more homo-genous cropping pattern) in Indian Punjab and Haryana wherethis study showed the closest ET to values given in literature,and much lower values in Rajasthan and Madhya Pradesh,where this study showed a significant lower ET.

The WP results show huge differences as expected for the largeheterogeneous IGB. The WP, by this study, agrees well withliterature values except in Madhya Pradesh, where Mishra et al.(1997) modeled point values with a significant high yield(Table 2). The WP values in Pakistani part of Indus basin arehigher than those given by Waterwatch (2003) which were alsobased on remote sensing. Further comparison revealed asignificantly higher ET by Waterwatch as a result of the expandedgrowth stage towards early June, when the pan evaporation was ata peak rate.

5. Discussions and conclusions

Rice yield, water consumption and WP are relatively low withtremendous variation in the IGB, indicating significant scope forimprovement. Overall, the WP values are more influenced by yieldin the range shown in Fig. 5(a). The WP values increase with theincrement of yield. The major challenge would be to increase yieldwhich will lead to higher ET. While yield increment usually takesyears to happen, another option to improve WP simultaneouslywould be to reduce ET of low-yield areas. It can be observed thatthe areas with low yield (below 3 tons/ha) have relatively morevariation in WP values, indicating greater differences in waterconsumption in spite of similar yields. The possible scope forimprovement (S2–S1) is to be achieved more likely in low-yieldareas than in high-yield areas. The yield relation to ET is morediverse as shown in Fig. 5(b). The ‘‘bright spot’’ of high WP areasaround Indian Punjab has much higher yield and WP values(cycled). It even creates significant changes on the slope ofyield � ET relation (S2–S3). This outstanding patch sheds light onthe way ahead for other under-performing areas which form themajor body of the basin.

Fig. 6 shows the crop water stress map as the ratio of rice ETa toETp along with the rainfall distribution map of the rice-growingperiod measured from Tropical Rainfall Measuring Mission(TRMM). It is observed that the rainfall is much lower in theIndus basin, which is opposite to ETp distribution. Significant cropstress (the lower the ratio the greater the stress) is observed in alarge part of the basin. The rice yield in IGB is generally lowcompared with that in many other parts of the world. Low yieldcould be attributed to many constraining factors including variety,

Table 2Comparison on yield, ET and water productivity of rice for the IGB from this study and literature.

Location ET (mm) Yield (ton/ha) Water productivity (kg/m3) References

This studya Literature This study Literature This study Literature

Punjab, India 534 N/Ab 6.18 N/A 1.18 0.4–0.5 Khepar et al. (1997)

408–527 4.6–6.4 0.87–1.46 Singh et al. (2001)

430–579 5.7–8 1.1–1.63 Mahajan et al. (2009)

Northwestern India 489 N/A 4.92 N/A 1 1.1 Sandhu et al. (1980)

N/A N/A 0.5–1.1 Tuong and Bouman (2003)

Madhya Pradesh, India 256 N/A 1.28 4.6–6.4 0.57 0.89 Mishra et al. (1997)

Sirsa, India 492 858–960 5.00 7.3–9 1.01 0.84–0.94 Singh et al. (2006)

500–600 1–5.5 0.7–098 Bastiaanssen et al. (2006)

Pindi Bhattian, Pakistan 472 544 2.69 2.8–3.7 0.57 0.4–1.6 Ahmad et al. (2004)

Indus basin, Pakistan 396 220–880 2.72 1–4 0.69 0.45 Waterwatch (2003)

Global 416 400–850 3.00 2.8–11 0.74 0.6–1.6 Zwart and Bastiaanssen (2004)

a WP values given for this study are averages of the division/state/basin as the coordinates of WP values from literature are unknown.b Data not available from the literature.

Fig. 5. Relations between (a) water productivity and yield and (b) yield and evapotranspiration of rice.

X.L. Cai, B.R. Sharma / Agricultural Water Management 97 (2010) 309–316314

Author's personal copy

soil, water and climate all of which lead to reduced transpiration.The most severe water stress occurs in the Madhya Pradesh andRajasthan states of India, where rainfall is lower and a large area israin-fed. Parts of Pakistani Sind and Punjab provinces have alsowitnessed significant water stress. The NDVI profiles showed thatthese areas have much more complicated cropping patterns incomparison to Indian Punjab. The diverse cropping types andgrowing periods could both lead to low average pixel ET. The well-performing Indian Punjab showed very little water stress.Surprisingly, large areas in the downstream of Ganges also showeda very low water-stress level, despite low yields and low WPvalues. In these areas the rainfall is very high, which iscomplemented by higher flow rates in the rivers. However, higherwater availability does not necessarily lead to higher yield or WP,as shown in this case. Crop development is linked to land, crop andwater management practices. Rainfall may occur at any time;hence the paddy has more standing water to evaporate but couldstill suffer from water stress during the critical crop growth period(especially the terminal grain filling stage) which drastically affectsthe amount of the final grain yield. Excess water itself could alsoimpose stress on rice growth at certain stages. Well-developedirrigation and drainage systems together with matching manage-ment practices can help maximize utilization of rainfall and riverflows to achieve high yield and WP. Other land and cropinterventions, such as laser land leveling, furrow-irrigated raisedbed (FIRBS) cultivation, insects and diseases control, fertilizer andvariety, are also important factors to be considered along withwater management.

This study developed a method to map large-scale yield, ET andWP of rice by integrating census, remote sensing and weather data,most of which are freely available online or from nationalauthorities. The extrapolation of yield data from district-levelstatistics to pixel-level values through NDVI bridges the publiclyaccepted official figures and advanced remotely sensed data. Themethod avoided complex land surface processes and biophysicalparameter estimations in remote sensing applications. It does notrequire field calibration prior to new applications and can be easilyapplied elsewhere. The accuracy is promising as long as censusproduction data fall within an acceptable range. The SimplifiedSurface Energy Balance model is another effort to bring simplicityin crop water consumption studies. Also based on the land surfaceenergy balance concept, but without complex parameter estima-tion processes, the model takes ET0 calculated from conventionalapproaches, such as P–M or Hargreaves’ equation, by using easilyaccessible data, multiplied with crop coefficient to calculate thepotential ET, which is then extrapolated to a large area based on thedistribution of land surface temperature. This kind of simple

structured method is ideal for large-scale operational monitoringof which automated processing is one of the key challenges. Theconventional approach to calculate potential ET of large scale alsotakes care of conditions such as altitude and air temperature,which could not be captured by remote sensing imagery. Theinherent strength of remote sensing, widely adapted authoritativecensus data and normal weather data are well-combined for large-scale WP studies. The method can be easily replicated in otherareas due to the simplicity of the process and the popularity of thedata set required.

There are a number of uncertainties in large-scale WP mapping.Quality of data is certainly an important issue to be considered.Census data are collected through a complicated countryprocedure through many labor inputs. Hence, errors and biasedvalues are likely to exist. Remote sensing data have to be convertedto ground values through many processes, such as sensorcalibration and atmospheric/topographic correction, during whichdistortions could remain. Crop type distribution is anotherimportant issue. To estimate crop production and water con-sumption one needs to know the exact location of the crops.However, mixed cropping patterns and fragmented farming arecommon practices in many agricultural systems, making it difficultto distinguish crop types over a large area. The sub-pixel mixturefurther complicates the situation. As shown in this study, the ricepixel ET tends to be lower than field-measured/modeled values,which could be largely attributed to the low ET values from otherland use types within the same pixel.

Acknowledgements

This study was supported by ‘Basin Focal Project for the Indo-Gangetic Basin’ under the CGIAR Challenge Program on Water andFood (CPWF). Detailed comments of two anonymous reviewers onan earlier version of the article greatly helped in providing clarityto the methodology and discussion of the results of the study.

References

Ahmad, M.D., Masih, I., Turral, H., 2004. Diagnostic analysis of spatial and temporalvariations in crop water productivity: a field scale analysis of the rice–wheatcropping system of Punjab, Pakistan. J. Appl. Irrig. Sci. 39 (1), 43–63.

Allen, R.G., Tasumi, M., Trezza, R., 2007. Satellite-based energy balance for mappingevapotranspiration with internalized calibration (METRIC)-model. ASCE J. Irrig.Drain. Eng. 133, 80–394.

Bastiaanssen, W.G.M., Ali, S., 2003. A new crop yield forecasting model based onsatellite measurements applied across the Indus basin, Pakistan. Agric. Ecosyst.Environ. 94 (3), 321–340.

Bastiaanssen, W.G.M., Menenti, M., Feddes, R.A., Holtslag, A.A., 1998. A remotesensing surface energy balance algorithm for land (SEBAL). 1. Formula. J. Hydrol.212–213, 198–212.

Fig. 6. TRMM rainfall during the rice-growing season and the ratio of rice actual ET to potential ET in IGB.

X.L. Cai, B.R. Sharma / Agricultural Water Management 97 (2010) 309–316 315

Author's personal copy

Bastiaanssen, W.G.M., Thiruvengadachari, S., Sakthivadivel, R., Molden, D., 1999.Satellite remote sensing for estimating productivities of land and water. Int. J.Water Resour. Dev. 15 (1/2), 181–196.

Bastiaanssen, W.G.M., Zwart, S.J., Pelgrum, H., 2006. Remote sensing analysis. In:Dam, J.C. van, Malik, R.S. (Eds.), Water productivity of irrigated crops in Sirsadistrict, India. WATRPRO Final Report. Wageningen, pp. 85–96.

Cai, X.L., Cui, Y.L., 2009. A simplified ET mapping algorithm and the application inZhanghe irrigation district. J. Irrig. Drain. 29 (in Chinese, with English abstract).

Cai, X.L., Sharma, B., 2009. A regional crop dominance map based on remote sensing,census and groundtruth data. In: Paper submitted to International Conferenceon Food Security and Environmental Sustainability, Kharagpur, India, Decem-ber 17–19.

Cai, X.L., Thenkabail, P.S., Biradar, C., Platonov, A., Gumma, M., Dheeravath, V.,Cohen, Y., Goldshlager, N., Ben-Dor, E., Alchanatis, V., Vithanage, J.V., Mar-kandu, A., 2009. Water productivity mapping using remote sensing dataof various resolutions to support ‘‘more crop per drop’’. J. Appl. Remote Sens.3, 033557.

Courault, D., Seguin, B., Olioso, A., 2005. Review on estimation of evapotranspirationfrom remote sensing data: from empirical to numerical modeling approaches.Irrig. Drain. Syst. 19 (3), 223–249.

Frolking, S., Qiu, J., Boles, S., Xiao, X., Liu, J., Zhuang, Y., Li, C., Qin, X., 2002.Combining remote sensing and ground census data to develop new maps ofthe distribution of rice agriculture in China. Global Biogeochem. Cycles 16 (4),1–10.

Frolking, S., Xiao, X., Zhuang, Y., Salas, W., Li, C., 1999. Agricultural land-use in China:a comparison of area estimates from ground-based census and satellite-borneremote sensing. Global Ecol. Biogeogr. 8 (5), 407–416.

Gonzalez, M.P., Mateos, L., 2008. Spectral vegetation indices for benchmarkingwater productivity of irrigated cotton and sugarbeet crops. Agric. Water Man-age. 95 (1), 48–58.

Groten, S.M.E., 1993. NDVI-crop monitoring and early yield assessment of BurkinaFaso. Int. J. Remote Sens. 14 (8), 1495–1515.

Harvey, J.T., 2002. Estimating census district populations from satellite imagery:some approaches and limitations. Int. J. Remote Sens. 23 (10), 2071–2095.

Hurtt, G.C., Rosentrater, L., Frolking, S., Moore III, B., 2001. Linking remote-sensingestimates of land cover and census statistics on land use to produce maps ofland use of the conterminous United States. Global Biogeochem. Cycles 15 (3),673–685.

Khepar, S.D., Sondhi, S.K., Kumar, S., Singh, K., 1997. Modelling effects of culturalpractices on water use in paddy fields—a case study. Research Bulletin, Pub-lication No. NP/SWE-1. Punjab Agricultural University, Ludhiana, India.

Loveland, T.R., Reed, B.C., Brown, J.F., Ohlen, D.O., Zhu, Z., Yang, L., Merchant, J.W.,2000. Development of a global land cover characteristics database and IGBPDISCover from 1 km AVHRR data. Int. J. Remote Sens. 21, 1303–1330.

Mahajan, G., Bharaj, T.S., Timsina, J., 2009. Yield and water productivity of rice asaffected by time of transplanting in Punjab, India. Agric. Water Manage. 96 (3),525–532.

McCoy, R.M., 2004. Field Methods in Remote Sensing. Guilford Press, New York,159 pp.

Mesev, V., 1998. The use of census data in urban image classification. Photogramm.Eng. Remote Sens. 64 (5), 431–438.

Mishra, H.S., Rathore, T.R., Pant, R.C., 1997. Root growth, water potential, and yieldof irrigated rice. Irrig. Sci. 17, 69–75.

Molden, D., Frenken, K., Barker, R., de Fraiture, C., Mati, B., Svendsen, M., Sadoff, C.,Finlayson, C.M., 2007. Trends in water and agricultural development. In:Molden, D. (Ed.), Water for Food, Water for Life: A Comprehensive Assessmentof Water Management in Agriculture. Earthscan, London, pp. 57–89.

Molden, D., Murray-Rust, H., Sakthivadivel, R., Makin, I., 2003. A water-productivityframework for understanding and action. In: Kijne, J.W., Barker, R., Molden, D.(Eds.), Water Productivity in Agriculture: Limits and Opportunities for Improve-ment. CAB International, Wallingford, UK, pp. 1–18.

Moulin, S., Bondeau, A., Delecolle, R., 1998. Combining agricultural crop models andsatellite observations: from field to regional scales. Int. J. Remote Sens. 19 (6),1021–1036.

Neto, E.R., Hamburger, D.S., 2009. Census data mining for land use classification. In:Cao, L., Yu, P.S., Zhang, C., Zhang, H. (Eds.), Data Mining for Business Applica-tions. Springer, USA, pp. 241–251.

Oza, S.R., Panigrahy, S., Parihar, J.S., 2008. Concurrent use of active and passivemicrowave remote sensing data for monitoring of rice crop. Int. J. Appl. EarthObs. Geoinf. 10 (3), 296–304.

Rindfuss, R.R., Stern, P.C., 1998. Linking remote sensing and social science: the needand the challenges. In: Liverman, D., Moran, E.F., Rindfuss, R.R., Stern, P.C.(Eds.), People and Pixels: Linking Remote Sensing and Social Science. NationalAcademy Press, Washington, DC, pp. 1–27.

Roerink, G.J., Su, Z., Menenti, M., 2000. S-SEBI: a simple remote sensing algorithm toestimate the surface energy balance. Phys. Chem. Earth B 25 (2), 147–157.

Rosegrant, M.W., Ringler, C., Gerpacio, R.V., 1997. Water and land resources andglobal food supply. In: Paper presented at the 23rd International Conference ofAgricultural Economists on Food Security, Diversification, and Resource Man-agement: Refocusing the Role of Agriculture, Sacramento, CA, August.

Sandhu, B.S., Khera, K.L., Prihar, S.S., Singh, B., 1980. Irrigation need and yield of riceon a sandy loam soil as affected by continuous and intermittent submergence.Indian J. Agric. Sci. 50 (6), 492–496.

Schmugge, T.J., Kustas, W.P., Ritchie, J.C., Jackson, T.J., Rango, A., 2002. Remotesensing in hydrology. Adv. Water Resour. 25, 1367–1385.

Senay, B.G., Budde, M., Verdin, J.P., Melesse, A.M., 2007. A coupled remote sensingand simplified surface energy balance approach to estimate actual evapotran-spiration from irrigated fields. Sensors 7 (6), 979–1000.

Shah, T., Singh, O.P., Mukherji, A., 2006. Groundwater irrigation and South Asianagriculture: empirical analysis from a large scale survey of India, Pakistan,Nepal and Bangladesh. Hydrogeol. J. 14 (3), 286–309.

Singh, K.B., Gajri, P.R., Arora, V.K., 2001. Modelling the effects of soil and watermanagement practices on the water balance and performance of rice. Agric.Water Manage. 49, 77–95.

Singh, R., Dam, J.C., van Feddes, R.A., 2006. Water productivity analysis of irrigatedcrops in Sirsa district, India. Agric. Water Manage. 82 (3), 253–278.

Sutton, P., Roberts, D., Elvidge, C., Baugh, K., 2001. Census from heaven: an estimateof the global human population using nighttime satellite imagery. Int. J. RemoteSens. 22, 3061–3076.

Thayyen, R.J., Gergan, J.T., Dobhal, D.P., 2005. Slope lapse rates of temperature in DinGad (Dokriani Glacier) catchment, Garhwal Himalaya, India. Bull. Glaciol. Res.22, 31–37.

Thenkabail, P.S., Biradar, C.M., Turral, H., Noojipady, P., Li, Y.J., Vithanage, J.,Dheeravath, V., Velpuri, M., Schull M., Cai, X.L., Dutta, R., 2006. An irrigatedarea map of the world (1999) derived from remote sensing. Research Report105. International Water Management Institute, Colombo, Sri Lanka, 74 pp.

Thenkabail, P.S., GangadharaRao, P., Biggs, T., Krishna, M., Turral, H., 2007. Spectralmatching techniques to determine historical land use/land cover (LULC) andirrigated areas using time-series AVHRR pathfinder datasets in the Krishna riverbasin, India. Photogramm. Eng. Remote Sens. 73 (9), 1029–1040.

Thiruvengadachari, S., Sakthivadivel, R. 1997. Satellite remote sensing for assess-ment of irrigation system performance: a case study in India. Research Report 9.International Irrigation Management Institute, Colombo, Sri Lanka, 23 pp.

Tuong, T.P., Bouman, B.A.M., 2003. Rice production in water scarce environments.In: Kijne, J.W., Barker, R., Molden, D. (Eds.), Water Productivity in Agriculture:Limits and Opportunities for Improvement. CAB International, Wallingford, UK,pp. 53–67.

Ullah, M.K., Habib, Z., Muhammad, S., 2001. Spatial distribution of reference andpotential evapotranspiration across the Indus basin irrigation systems. WorkingPaper 24. International Water Management Institute, Lahore, Pakistan.

Waterwatch, 2003. Crop and water productivity of the Pakistan wheat and ricesystems. Technical Report. Wageningen, The Netherlands.

Xiao, X., Boles, S., Frolking, S., Li, C., Babu, J.Y., Salas, W., Moore, B., 2006. Mappingpaddy rice agriculture in South and Southeast Asia using multi-temporal MODISimages. Remote Sens. Environ. 2006 (10), 95–113.

Yang, P., Wu, W., Zhou, Q., Zha, Y., 2008. Research progress in crop yield estimationmodels based on spectral reflectance data. Trans. CSAE 24 (10), 262–268 (inChinese with English abstract).

Zwart, S.J., Bastiaanssen, W.G.M., 2004. Review of measured crop water productivityvalues for irrigated wheat, rice, cotton and maize. Agric. Water Manage. 69 (2),115–133.

Zwart, S.J., Bastiaanssen, W.G.M., 2007. SEBAL for detecting spatial variation ofwater productivity and scope for improvement in eight irrigated wheat sys-tems. Agric. Water Manage. 89 (3), 287–296.

X.L. Cai, B.R. Sharma / Agricultural Water Management 97 (2010) 309–316316