Projection of net irrigation requirement over India using a bias-corrected spatially downscaled dataset

The net irrigation requirement (NIR) is the amount of water required to bring the soil in the crop root zone to field capacity at the time of irrigation. It is the measure of water essential for the growth of crops and differs on the cropping pattern and the growing climate. It can be determined by direct measurements of soil moisture, such as by using tensiometers, or indirect measurements of soil moisture loss, such as by estimating evapotranspiration (ET). Allen et al. (1998) defined ET₀ as ‘the rate of ET from a hypothetical crop with an assumed crop height (0.12 m) and a fixed canopy resistance (70 s/m) and albedo (0.23) which would closely resemble ET from an extensive surface of green grass cover of uniform height, actively growing, completely shading the ground and not short of water’. It is a key component in hydrological studies. It also plays a substantial role in regional and global climates through its contribution to hydrological cycles with more than 70% of rainwater returning to the atmosphere through ET (Ajami 2021). In the recent past, there has been an increase in the percentage of groundwater utilized for irrigation in India. According to the Central Ground Water Board (CGWB 2020), over 60% of the total area irrigated in the country is being met by groundwater resources. The increasing dependency on groundwater as a source of irrigation water without proper recharging capacities of groundwater leads to the depletion of groundwater and lowers the water table. This creates an urgent need for the management of groundwater resources. However, due to considerable variations in the development of groundwater in the country, the management of groundwater resources is extremely complex (Jha & Sinha 2009). Therefore, any strategy involved in the management of groundwater resources should consider both the supply and demand side at the regional scale. One such important scientific strategy lies in controlling the water demand in various sectors, especially in the agriculture sector, which accounts for 92% of the groundwater withdrawal each year (Jha & Sinha 2009). Therefore, accurate estimation of ET₀ is highly important in analysing groundwater recharge, land-use planning, prediction of crop yields, etc. Estimating ET by direct measurements (using a Lysimeter) or indirect measurements of soil moisture can help in finding out the NIR. But this is a difficult and tedious process. Therefore, a general practice is employed where the ET₀ is first estimated from a standard surface and then an appropriate empirical crop coefficient (k_c) prepared by FAO is applied for each crop, which accounts for the difference between the ET₀ and crop ET (ET_c) (Allen et al. 1994). The use of the ET₀-k_c approach has been immensely effective in avoiding the necessity for calibrating the ET equation separately for each crop and its stage of growth (Wright 1982; Jensen et al. 1990). It, moreover, provided a working model which can be used until more sophisticated methods become available for direct estimation of actual ET_c (Bandyopadhyay et al. 2012).

However, the growing conditions of actual vegetation in fields are questionable in comparison to the reference conditions representing the ideal k_c value (Allen et al. 2007), especially in water shortage areas. Moreover, it is difficult to obtain the correct information on the types of crops and their characteristics for large populations of crops and fields. In such cases, satellite data becomes an ideal option to generate spatially continuous ET. Satellite-based technologies are further found to be highly reliable and accurate in estimating ET whilst providing information on large areas with minimum cost and time (Tasumi 2019). Researchers have developed many such models that employ satellite datasets for calculating the instantaneous ET as a residual of the energy balance at the Earth's surface. Such surface energy balance models include Surface Energy Balance Algorithm for Land (SEBAL) (Bastiaanssen et al. 1998), mapping ET at high resolution with internalized calibration (METRIC) (Allen et al. 2007), Surface Energy Balance Systems (SEBS) (Su 2002), and operational simplified surface energy balance (SSEBop) (Senay et al. 2013), etc. The main advantage of ET estimation using surface energy balance models over conventional methods is that neither the crop grown nor its developmental stages need to be known (Allen et al. 2007). Another advantage is that energy balance models can detect a reduction in ET due to a shortage of water available in the area. It is also found that remote sensing-based ET estimation gives accurate data when dealing with larger scales as compared to point-based estimation (Wagle et al. 2017). The present study tries to understand the general idea of adopting satellite data to generate the k_c layer. Further, with the vast study area (India, 3.29 M km²), collecting information on the specific type of crop grown and other characteristics such as development stages, nature of growth, water availability, etc. of each cropping area is impossible. Therefore, we hypothesized that the k_c layer generated by using surface energy balance models, which employ satellite images, could essentially be used for adequately estimating the NIR of the study area.

With changing climatic conditions and a growing population, future water availability is expected to further decline, raising demands for higher crop productivity (Brauman et al. 2013). Further, the most common approach to study hydrological response to the change in climate is to run hydrological models using projected climatic variables (precipitation and temperature) from global climate models under different climate scenarios, downscaled or bias-corrected to the considered study area (Fowler et al. 2007; Chiew et al. 2009; Senatore et al. 2011). The National Aeronautics and Space Administration (NASA)'s Earth Exchange Global Daily Downscaled Projections (NEX-GDDP) datasets consist of downscaled climatic scenarios that are available globally. They are derived from the General Circulation Model (GCM) runs executed beneath the Coupled Model Intercomparison Project Phase 5 (CMIP5) and across two Greenhouse Gas (GHG) emissions, which are generally known as RCP. The NEX-GDDP dataset provides support to the science community in conducting investigations regarding the effects of the change in global climatic conditions at local to regional scales. They also attempt to improve public awareness of the probable future climatic patterns spatially at the local or regional level. The NEX-GDDP dataset incorporates downscaled daily scenarios that are projected from 21 models circulated under CMIP5.

In this study, an attempt was made to estimate NIR for three different cropping pixels, i.e., Kharif-only, Rabi-only, and double/triple using satellite-derived monthly k_c layers for the present climate. Further, using the NEX-GDDP future climatic projections, the probable relative percentage changes in the ET₀ and NIR for the projected climate were estimated. The present study has been taken up with the following objectives:

• i.

  To calculate ET₀ from the IMD gridded temperature dataset,

• ii.

  To prepare satellite-based monthly k_{c_act} layers and develop monthly NIR maps for the present climate,

• iii.

  To estimate ET₀ and the corresponding NIR for the future projected climate.

India, the 7th largest country in the world with an area of 3,287,263 km² comprising 90.08% land area and 9.92% water, was selected as the study area. The country is situated north of the equator between 8°4′ and 37°6′ N latitudes and between 68°7′ and 97°25′ E longitudes. The country experienced four seasons annually, namely, summer, monsoon, post-monsoon, and winter. The summer season in India starts in the month of March and is extended until May, followed by the monsoon (rainy) season, which starts from June to September. The post-monsoon period extends from the month of October to December and lastly, the winter season occurs during January and February. Temperatures can exceed 40 °C during the day during the summer season in India, with 50.6 °C as the highest temperature recorded in Alwar, Rajasthan, during the summer of 1956. Summers in India are usually extended until June. In the year 1895, a temperature of −20 °C was recorded in Srinagar, Jammu and Kashmir, which was the lowest temperature recorded in India as per the India Meteorological Department (IMD) until 2010. India also gets high rainfall during the monsoons. The 24 h heaviest rainfall was received in Cherrapunji, Meghalaya, which was 1,563 mm in the year 1995 (IMD).

Severe water shortages are developing in India and agricultural water is becoming increasingly scarce due to the increasing demand for water from different sectors. Groundwater exploration for irrigation in India plays a major role in increasing food production (Ganesh Babu et al. 2014). However, the over-exploitation of groundwater in many states of the country (especially in Punjab, Haryana, and Rajasthan) necessitates the understanding of the crop water requirement for different crops. The consumption of water by the agriculture sector in India is as high as 81%, implying that efficient use of water in agriculture is our topmost priority (Surendran et al. 2013). It is a known truth that water supply sources for irrigation are important for agricultural production and this current increase in irrigation water use is considered unsustainable and threatens food production. NIR, if known accurately, can improve the economy of a country by judiciously applying the required amount of water in appropriate intervals of time and can thus help irrigation scheduling and overall improve crop productivity. This will not only help in the conservation of groundwater but also to accomplish effective irrigation management.

India has mainly three cropping seasons, namely, Kharif, Rabi, and Zaid. The Kharif season in India usually starts in the month of June and ends in October, while the Rabi season starts in November and is extended until March. Zaid is also known as a summer crop, which is usually grown between the month of March to May. The present study deals with the estimation of NIR for three cropping patterns, namely, Kharif, Rabi, and double/triple-cropping patterns. The double/triple-cropping pattern essentially refers to that area where there is continuous cultivation of crops throughout the year (Kharif–Rabi–Zaid). In this case, all 12 months were considered as the cultivated period, which consists of three cropping seasons, namely, Kharif, Rabi, and summer crops (Zaid).

The newly developed daily rainfall dataset (Pai et al. 2014) available at a spatial resolution of 0.25° × 0.25° for the period 1971–2010 was used in this study (Table 1). A total of 3,964 grid points covered the entire study area. Further, daily gridded maximum and minimum temperature datasets with a spatial resolution of 1° × 1° covering the entire country developed by Srivastava et al. (2009) were employed for the present study. A total of 362 grid points covered the whole of India. Bandyopadhyay et al. (2018) compared these two datasets along with other reanalyses datasets with IMD point data and found them to be suitable for further analysis.

The Shuttle Radar Topography Mission (SRTM) Digital elevation model (DEM), produced by NASA originally, is a major invention in digital mapping of the world and has provided a major improvement in the availability of high-quality elevation information for huge portions of the tropics and other areas. The entire 90 m × 90 m resolution SRTM-DEM layer for the entire study area was downloaded. The LULC raster layer of the entire country with a spatial resolution of 5 km × 5 km was obtained from the Space Application Centre (SAC), Ahmedabad. The entire study area is classified under 25 different land-use classes, namely, build-up, current fallow, deciduous forest, double/triple, evergreen forest, grassland, gullied, hill slope forest, hill slope other than forest, irrigated area other than rice, Kharif-only, littoral swamp, other wastelands, plantation/orchard, Rabi-only, ram, rice, scrub/degraded forest, scrubland, shifting cultivation, snow, water bodies, and Zaid only. Among these, only three cropping pixels, namely, Kharif-only, Rabi-only, and double/triple-cropping pixels were masked out for the present study.

The ET_r and ET₀ can be a daily or monthly value, or they can also be derived from gridded data collected from gridded weather data. The cloud-free images of the ET₀F layer, which was considered as the k_c layer was downloaded for the entire study area for 6 years, i.e., 2001–2006. Since the cloud-free layers were much less, images with a cloud coverage of up to 10% were also considered for the study. The characteristics of the ET₀F layer are mentioned in Table 1.

The FEWSNET data centre offers access to substantial data associated with food security, including their classifications, administrative boundaries, livelihood zone, remote sensing imagery, prices, and trade carried out at the borders. For the present study, the monthly ET_a and the daily potential ET (ET_{0_FEWSNET}) were downloaded. Their details are discussed below.

The monthly ET_a provided by FEWSNET is produced using the operational Simplified Surface Energy Balance (SSEBop) model (Senay et al. 2013) from 2003 to the present day. The SSEBop model combines the fraction of ET that is produced from remotely sensed Moderate Resolution Imaging Spectroradiometer (MODIS) thermal imagery, obtained every 10 days, with reference ET using the thermal index approach. Compared to other surface energy models, the salient feature of the SSEBop model is that it employs pre-defined, seasonally dynamic boundary conditions that are unique to individual pixels for the ‘hot/dry’ and ‘cold/wet’ reference points. The original formulation of SSEB is based on manual selection of the ‘hot/dry’ and ‘cold/wet’ pixels, which is the basic ideology of SEBAL (Bastiaanssen et al. 1998) and METRIC (Allen et al. 2007) models. The FEWSNET ET_a data can be downloaded from https://earlywarning.usgs.gov/fews/product/460. The data obtained have a spatial resolution of 0.25° × 0.25° for global extent and are available as Tagged Image File Format (TIFF). For the present study, the monthly ET_a was downloaded for 8 years, from 2003 to 2010.

Similarly, the daily ET_{0_FEWSNET} product of FEWSNET was also acquired for the present study. The daily ET_{0_FEWSNET} is estimated from climatic parameters, which are obtained from Global Data Assimilation System (GDAS) analysis fields. ET_{0_FEWSNET} was calculated spatially using the equation for estimation of reference crop evaporation in accordance with FAO 56, explained in Allen et al. (1998). This essentially indicated that the ET_{0_FEWSNET} layer produced is a crop reference ET (ET₀). The daily ET_{0_FEWSNET} layer has a spatial resolution of 1° × 1° and is global in its spatial extent. For the present study, the daily ET_{0_FEWSNET} was downloaded for 8 years from 2003 to 2010. The ET_{0_FEWSNET} data can be downloaded from https://earlywarning.usgs.gov/fews/product/81.

The NASA Earth Exchange Global Daily Downscaled Projection (NEX-GDDP) precipitation and temperature datasets comprises globally downscaled and bias-corrected climatic scenarios that are a derivative of the Global Climate Model (GCM) runs executed beneath the CMIP5 and across two of the four GHG emission scenarios known as Representative Concentration Pathways (RCPs). The NEX-GDDP dataset includes downscaled projections for RCPs 4.5 and 8.5 scenarios from 21 CMIP5 GCMs for which daily climatic scenarios were prepared and distributed under CMIP5. Each climatic projection incorporates daily maximum and minimum temperatures, and precipitation for the periods from 1950 to 2005 (‘Retrospective Run’) and from 2006 to 2100 (‘Prospective Run’). All the datasets obtained from NEX-GDDP have a 0.25° × 0.25° spatial resolution and are global in spatial extent.

The ET₀F layer, which is considered as the traditional k_c was obtained for 6 years from 2001 to 2006. While downloading these layers, precautions were taken such that only cloud-free layers were chosen for the study because it might hamper the quality of the data being retrieved. However, because of the shortage in the availability of cloud-free layers, images with cloud coverage of up to 10% were also considered. In addition, due to fewer cloud-free images, an attempt was made to obtain at least one cloud-free image monthly to represent the monthly layer, although it was difficult in some areas, especially during the months of July and August. Further, due to the vast study area, the acquisition period of all images is not the same. Therefore, images acquired closest to the middle of the month were chosen with preferences given to those days without any cloud cover. The data downloaded were from Landsat 5 and 7 based on calibrated images. Another problem encountered with the data was the Landsat 7 Enhanced Thematic Mapper Plus (ETM+) Scan Line Corrector (SLC)-off which showed gaps in the images. However, this problem could be easily solved as the ET₀F value of those pixels under these gaps was given ‘zero’. Therefore, any pixels which have zero or less were set as a Null/NoData value. Also, some outliers containing very high values were also encountered, which were concurrently removed using a ‘SET NULL’ function of ArcMap. These outliers in ET₀F might be because of the systematic error caused by several assumptions made in the energy balance model (Foolad et al. 2018). For filling in the gap (Null pixels), the mean of the surrounding pixels was taken while ignoring the Null value and the images were then resampled to match with the LULC layer.

Another satellite-based monthly k_c layer was generated by taking the ratio of the FEWSNET-provided monthly ET_a (using the SSEBop model) to the monthly ET_{0_FEWSNET} layers (Penman Monteith crop reference estimation method), i.e., ET_a/ET_{0_FEWSNET} (Allen et al. 2007). The downloaded raster layers had some unknown/missing pixels, which were first set as Null before being processed. Since the ET_{0_FEWSNET} layers obtained were daily rasters, they were converted into monthly rasters by using ArcMap. Further, since the daily ET_{0_FEWSNET} layers were scaled up by 100 to maintain the accuracy to 0.01 mm, the monthly ET_{0_FEWSNET} layers were afterwards divided by 100.

The NASA NEX-GDDP daily data of precipitation and maximum and minimum temperatures were downloaded from 1979 to 2005 as historical data (baseline) and from 2019 to 2099 as future projected scenarios. The downloaded data were in .nc format, which were converted to TIFF using QGIS software. Due to the vast study area and a huge number of sub-layers (daily), it was impossible to directly clip off for the study area. Therefore, the latitude and longitude of all the grid points obtained from IMD precipitation data (0.25° × 0.25°) were employed to generate the time series of the points covering the entire country. Point shapefiles were made for each monthly time series data covering the entire study area. The shapefiles were then converted to a raster layer, keeping the cell size as 0.25° × 0.25°. The layers were resampled to match the LULC layer for further analysis.

The bias-corrected NEX-GDDP daily precipitation and maximum and minimum temperatures for the present and future years for all the four models selected for the present study, i.e., CCSM, CNRM, MPI, and MRI were analysed to determine the relative and absolute change in the future when compared with the present climatic scenario. The downscaled precipitation and temperature data from 1979 to 2099 were divided into four time slices. The period from 1979 to 2005 is considered the historic timeline, which is taken as the baseline for the present study. The future projected period from 2019 to 2099 was divided into three time slices, i.e., near-future/2030s (2019–2045), mid-future/2060s (2046–2072), and far-future/2080s (2073–2099) under RCPs 4.5 and 8.5 scenarios containing 27 years each. The calculated monthly maximum and minimum temperatures and precipitation data for baseline (1979–2005) and projected future years (2019–2099) were used as climate forcings in the ‘ET₀ and NIR toolbar’ to estimate the NIR, ET_cm, and R_e. The monthly spatial mean of NIR, ET_cm, and R_e was estimated for the three cropping pixels (Kharif, Rabi, and double/triple) for baseline and all the future time slices under RCPs 4.5 and 8.5 for all the four models considered and their ensemble.

Further, an attempt was made to check the decadal variation of ET₀. For this purpose, the time taken for observation, i.e., 1971–2010 was divided into four decades, i.e., 1971–1980, 1981–1990, 1991–2000, and 2001–2010 and ET₀ was estimated for each decade. The annual average of ET₀ was maximum during the period 2001–2010 (5.25 mm/day) and the minimum was reported during 1971–1980 (5.19 mm/day). However, the ET₀ during the Kharif season was observed at its maximum during 1981–1990 (4.89 mm/day), while the maximum ET₀ during the Rabi season was observed during 2001–2010 (5.26 mm/day). This suggested the increasing ET₀ during the winter season, which might be due to a greater rise in temperature during the winter months rather than the summer months, as observed in the trend analysis of temperature, where the maximum increasing trend was found in the month of February, covering nearly 54% of the country (Nengzouzam et al. 2019). Chakravarty et al. (2015) also concluded similar observations, where they also reported an increase in ET₀ during the Rabi season. Comparing the ET₀ during the first decade of observation, i.e., 1971–1980 and the last decade, i.e., 2001–2010, the relative increase in ET₀ was observed to be higher during the Rabi season than the Kharif season by 1.44 and 1.20%, respectively.

Assuming the prepared k_{c_EEFlux} layer is similar for the whole period considered in the study, i.e., 1971–2010, the ET_cm, R_e, and NIR were estimated using ‘ET₀ and NIR toolbar’ for three different cropping pixels, i.e., Kharif-only, Rabi-only, and double/triple-cropping pixels.

The seasonal ET_cm, R_e, and NIR were also estimated for Rabi-only pixels. The ET_cm during this cropping season was found to be highest in the month of December with a spatial mean of 149.66 mm, while it was found to be lowest in the month of November with a spatial mean of 137.22 mm. The spatial average ET_cm of Rabi-only pixels was reported to be 713.53 mm for the Rabi season, which was found to be much higher than that of the Kharif season, which was unexpected. The reason for this might be that the EEFlux is struggling to account for background evaporation at the hot pixel calibration point, as hot pixels assigned have a huge impact on ET and are thereby impacted by errors or biases in the overall surface energy balance (Foolad et al. 2018). Another likely reason might be the introduction of biases during the application of EF to extrapolate instantaneous ET_rF to daily ET_rF (Foolad et al. 2018). Further, the aridity biases in the gridded weather data used for calculating the ET_r by EEFlux might not be as accurate for the area considered in the present study. A similar discrepancy was also reported by Filgueiras et al. (2019), who found a higher estimate of ET when compared with the SEBAL model for areas with lower NDVI and higher surface temperature. Ayyad et al. (2019) also reported that the EEFlux ET_a overestimates and deviates significantly from the SEBS-adjusted model by 36.36% and concluded that the USGS-FEWSNET SSEBop product gave the best performance. Due to the very low rainfall received during these cropping months, the total estimated average R_e during this season was reported to be as low as 57.91 mm. The R_e was found to be highest during the month of November with a spatial mean of 29.82 mm, which corresponded to the minimum NIR reported during this month with a spatial mean of 3.64 mm/day. The NIR was reported to be the highest in the month of February with a spatial mean of 4.58 mm/day. Similarly, the seasonal NIR for the Rabi-only pixels reported a spatial mean of 4.32 mm/day with a total of 648.74 mm (Figure 6(a)).

The annual ET_cm, R_e, and NIR estimated for double/triple-cropping pixels reported a similar trend as seen under Kharif and Rabi-only pixels. High ET_cm values stretch from September to March, while it was found to be lower during the months of April to June, which was unexpected. High NIR was reported from October to March, which might be due to the very low rainfall received during these periods (Maza et al. 2020). The annual ET_cm, R_e, and NIR reported ET_cm were reported to be 1,537.60, 1,037.00, and 1,036.8 mm, respectively (Figure 6(a)).

Due to the unexpectedly higher ET_cm with k_{c_EEFlux} during the Rabi season, an attempt was made to prepare another k_c layer using the ratio of the FEWSNET monthly ET_a and ET_{0_FEWSNET} for 8 years (2003–2010) as the monthly k_c layer.

The ET_cm, R_e, and NIR were also estimated for Rabi-only pixels. The ET_cm was found to be highest in the month of February (103.2 mm) and lowest in December (87.01 mm) (also visible in Bhadra et al. (2012)). The average monthly ET_cm estimated using k_{c_FEWS} followed a similar trend throughout compared to Bhadra et al. (2012), although the estimated ET_cm was reported to be lower during the summer months. This might essentially be due to the difference in the method with which the k_c and ET₀ layers are obtained. Another likely reason might be the lower ET₀ estimates during these months as HS tends to underestimate under lower RH (Valle Júnior et al. 2021).

To further check the accuracy of the estimated ET_cm and NIR, a few stations were selected, whose annual and seasonal ET_cm and NIR were compared with conclusions made by other researchers. The average monthly ET_cm and NIR estimated using k_{c_FEWS} followed a similar trend throughout compared to Bhadra et al. (2012), whereas the opposite trend was observed with the ET_cm estimated using k_{c_EEFlux}. Both the estimated average ET_cm using k_{c_FEWS} and k_{c_EEFlux} for the Kharif season for the Limbasi command area in Gujarat were found to be closely matching with that of Khandelwal & Dhiman (2017), while they were overestimated for the Rabi season using k_{c_EEFlux}. Similar observations were made when the estimated ET_cm and NIR were compared for the cultivated area of Anantapur (Ganesh Babu et al. 2014), such that k_{c_EEFlux} was highly overestimated during the Rabi season.

Further, to estimate the total volume of the irrigation water demand seasonally and annually, a summation of the NIR obtained from ‘ET₀ and NIR toolbar’ multiplied with the area covered by each pixel was carried out for each of the three cropping pixels employed in the study (Page et al. 2020). In doing so, the seasonal NIRs for the Kharif-only, Rabi-only, and double/triple-cropping pixels were estimated to be 60,740.67, 49,936.44, and 1,48,709.58 Mm³ considering a 100% irrigated area. This indicated that an average of 259.39 BCM (billion cubic metres) of water is required for irrigating Kharif-only, Rabi-only, and double/triple-cropping pixels altogether each year. On average, India stores only 8% out of the total 4,000 BCM of precipitation received every year (IWRS, https://iwrs.org.in/indias-water-resources/) out of which 3000 BCM are received during the monsoon season (Asian Development Research Institute, https://www.adriindia.org/adri/india_water_facts). While uniformity in the distribution of precipitation in the country is a concern, it is highly variable with less than 100 mm of annual rainfall received in Rajasthan to greater than 2,500 mm in Assam (Central Ground Water Board 2014). To top it all, more than 80% of stored water resources are utilized by the agriculture sector (Surendran et al. 2013). This strongly suggested the huge dependency of groundwater as a source of irrigation water in many parts of the country. While the volumetric NIR was estimated with the assumption of 100% irrigated area for the cropped pixels, the actual cultivated area under irrigation was unavailable for this study. Moreover, the dynamic nature of cultivated areas (cropped pixels) and most importantly, those brought under irrigation, which depends on the increasing food demand of the growing population, will have a considerable impact on the volumetric irrigation water requirement in the future. But due to the unavailability of sufficient data to project this change in irrigated areas, the present study was confined to NIR estimations in terms of depth only.

Regional-wise, the centrally located states such as Uttar Pradesh and Madhya Pradesh, which are the leading states in terms of agriculture, reported the highest NIR among others in all three cropping pixels (Figures 8, 10, and 12) especially during the Kharif season. States like Haryana, Maharashtra, Rajasthan, and Andhra Pradesh also showed higher NIR among the rest. The above-mentioned states are known to be one of the highest producers of agriculture products in the country, however, they receive the lowest rainfall. According to the CGWB (CGWB 2020), the stage of groundwater development exceeds 100% in states like Punjab (164.42%), Rajasthan (150.22%), and Haryana (134.56%) and is more than 50% in many other agriculture-producing states such as Uttar Pradesh, Madhya Pradesh, Maharashtra, Gujarat, etc. This implies that the groundwater extraction is much higher compared to the total annual groundwater recharged in some states, while it is depleting in many others. It also suggested that most agriculture-producing states are highly dependent on groundwater for irrigation. In this regard, an efficient use of water resources should be our top priority at this point. Judicious use of water resources for domestic, industrial, and irrigation purposes should be our major concern. Efficient irrigation practices such as drip irrigation should be practiced as far as possible in cultivated areas.

To check the reliability of these models, three performance indicators, namely NSE, CRM, and R² were employed and the ensemble of all the selected models was compared with the IMD station data. While comparing the precipitation data, out of 12 stations, 6 stations showed NSE greater than 0.5, 8 stations showed R² greater than 0.6, and 11 stations showed CRM values between −0.5 and 0.5, suggesting a satisfactory relation between the models and the station data (Table 2). The performance of the maximum temperature showed that all stations were found to secure NSE greater than 0.5 out of which 10 stations showed an NSE greater than 0.8, suggesting a near-to-perfect match between the model and station data. Further, all the stations showed R² greater than 0.8 and the CRM values were all in the range of −0.2 to 0.2, suggesting a very close match between the two. Similarly, the ensemble monthly minimum temperature of NASA NEX-GDDP was also compared for 12 stations with the IMD station data, where all stations reported NSE values greater than 0.7 out of which 8 stations were found to have NSE values greater than 0.9. Further, all the stations had R² values greater than 0.8 out of which 11 stations have R² values greater than 0.9, suggesting a very good correlation between the model and station data. All 12 stations showed CRM values in the range of −0.2 to 0.2, suggesting a near-to-perfect match.

The monthly input climate parameters obtained from NEX-GDDP for the present climate (baseline period) and future years for both RCPs 4.5 and 8.5 were analysed to determine the future changes in input climatic parameters (temperature and precipitation). However, since we are interested in three cropping pixels only, i.e., Kharif-only, Rabi-only, and double/triple, the monthly estimates were consolidated to obtain seasonal (Kharif and Rabi) and annual climate change.

Under both RCPs 4.5 and 8.5, the ensemble showed an increasing trend throughout the future time slices, such that maximum relative change was observed in the far-future (2080s). For RCP 4.5, a relative change of 12.48, 15.81, and 13.16% was reported for Kharif, Rabi, and annual, over the baseline averages of 835.51, 60.44, and 977.86 mm, respectively (Table 3). Whereas, for RCP 8.5, an increase in rainfall of 22.85, 28.76, and 22.77% was observed for Kharif, Rabi, and annually, respectively, over baseline. Chaturvedi et al. (2012) also projected an increase in precipitation from up to 14% under RCP8.5. A Ministry of Earth Sciences report (MoES report 2020) also predicts an increase in precipitation under both RCP 4.5 (6.0–18.2%) and RCP 8.5 (10.1–22.5) over India. Among the models considered, MRI reported the maximum increase in rainfall with as much as 34.55% annually (Table 3), while MPI reported the least (8.82% annually over baseline). Similarly, Choudhury et al. (2022) also projected an increase in rainfall (10.1–22.5%) under RCP 8.5 for a watershed in the northeastern hill region.

Under both RCPs 4.5 and 8.5, the ensemble showed rising temperature throughout the future time slices such that the maximum absolute increase in maximum temperature was observed in the far-future (the 2080s). For RCP 4.5, an absolute increase in maximum temperature of 1.72, 2.21, and 1.98 °C was reported for Kharif, Rabi, and the annual, respectively, over the baseline average of 30.56, 25.52, and 29.07 °C, respectively, in the 2080s (Table 4). Whereas, for RCP 8.5, an increase in maximum temperature of 3.28, 3.96, and 3.66 °C was observed for Kharif, Rabi, and annually, respectively, over baseline, in the 2080s. MPI reported the highest increase in the maximum temperature with as high as 4.69 °C annually (in the 2080s of RCP 8.5), while CNRM reported the least (2.66 °C annually) (Table 4).

Similarly, the maximum absolute increase in the minimum temperature was also reported in the 2080s under RCP 8.5. In the 2080s under RCP 4.5, an absolute increase in maximum temperature of 1.86, 2.39, and 2.14 °C was reported for Kharif, Rabi, and annual, respectively, over baseline averages of 21.22, 11.68, and 17.20 °C (Table 5). Whereas in the 2080s under RCP 8.5, an increase in maximum temperature of 3.63, 4.49, and 4.09 °C was observed for Kharif, Rabi, and annually, respectively. This showed a definite higher temperature rise in the minimum temperature as compared to the maximum temperature. These results matched well with the findings of Bandyopadhyay et al. (2009), who also reported more increase in the minimum temperature than the maximum temperature in observed data during 1971–2002. Similar results were also observed by Lobell et al. (2007), who observed a slightly greater increase in minimum temperature (2.4 °C during 2046–2065) compared to the maximum temperature (1.9 °C during 2046–2065) over India. Among the models considered, MPI reported the highest increase in the maximum temperature with as high as 4.97 °C annually (in the 2080s of RCP 8.5), while CNRM reported the least (3.52 °C annually) (Table 5).

The monthly ET₀, ET_cm, R_e, and NIR were estimated for the projected precipitation data for all four time slices, i.e., baseline (1979–2005), near-future or 2030s (2019–2045), mid-future or 2060s (2046–2072), and the far-future or 2080s (2073–2099) under RCPs 4.5 and 8.5 using k_{c_FEWS}. However, since we are interested in three types of cropping pixels only, the monthly ET_0, ET_cm, R_e, and NIR were carried out for Kharif-only (June to October), Rabi-only (November to March), and double/triple (all months). The monthly estimates were then consolidated to obtain seasonal and annual relative change over the baseline.

The spatial average ET_cm under RCPs 4.5 and 8.5 for Kharif, Rabi, and double/triple-cropping pixels showed a rising trend over the ensemble baseline throughout the future time slices such that the maximum relative change was observed in the far-future (2080s) (Table 7, Figure 13(b)). Under RCP 4.5, the ensemble reported the highest relative changes of 2.68, 5.51, and 4.00% over the ensemble baseline of 612.68, 408.43, and 1,263.25 mm, for Kharif, Rabi, and double/triple-cropping pixels, respectively, in the 2080s (Table 7). Whereas in the 2080s under RCP 8.5, an increase in ET_cm of 5.50, 9.25, and 7.00% over the baseline was observed for Kharif, Rabi, and double/triple-cropping pixels, respectively. Our findings indicated that the increase in ET_cm during the Rabi season is much higher than that of the Kharif season, corresponding to the greater rise in temperature during the Rabi season. The results also indicated the much higher change occurring under RCP 8.5, which is nearly twice the change occurring in RCP 4.5 both seasonally and annually (Table 7). MPI reported the maximum increase in ET_cm with as high as 10.4% annually (in the 2080s of RCP 8.5), while CNRM reported the least (2.71%) (Table 7).

The spatial average R_e under RCPs 4.5 and 8.5 for Kharif, Rabi, and double/triple-cropping pixels showed a rising trend over the ensemble baseline throughout the future time slices such that the maximum relative change was observed in the far-future (2080s) (Table 8, Figure 13(c)). In the 2080s under RCP 4.5, the ensemble reported the highest relative change of 15.41, 12.36, and 14.51% over the ensemble baseline of 525.92, 22.94, and 570.43 mm, for Kharif, Rabi, and double/triple-cropping pixels, in the 2080s (Table 8). Whereas in the 2080s under RCP 8.5, an increase in R_e of 22.54, 21.60, and 23.67% over the baseline was observed for Kharif, Rabi, and double/triple-cropping pixels, respectively. Corresponding to the maximum increase in precipitation reported by MRI, we observed the highest R_e with as high as 33.28% annually (in the 2080s of RCP 8.5), while MPI reported the least (8.37%) (Table 8). The relative percentage change of the ensemble seasonal and annual R_e under both RCPs 4.5 and 8.5 indicated a greater change occurring under RCP 8.5, which is nearly twice the change occurring under RCP 4.5 seasonally.

The spatial average NIR under RCPs 4.5 and 8.5 for Kharif, Rabi, and double/triple-cropping pixels showed an increasing trend over the ensemble baseline throughout the future time slices such that the maximum NIR over India was observed in the far-future (in the 2080s) (Table 9, Figure 13(d)). For Kharif-only pixels (during the Kharif season), the ensemble NIR showed a negative relative change corresponding to the increasing rainfall throughout the future time slices, under both RCP 4.5 (−14.47% in the 2020s to −7.25% in the 2080s over baseline of 1.2 mm/day) and RCP 8.5 (−11.45% in the 2020s to −5.59% in the 2080s) (Table 9). However, for Rabi-only pixels (during the Rabi season), even though a greater increase in rainfall was reported, a positive relative change in NIR was visible throughout the future time slices under both RCP 4.5 (3.30% in the 2020s to 5.21% in the 2080s over baseline NIR of 2.55 mm/day) and RCP 8.5 (2.34% in the 2020s to 8.83% in the 2080s). The spatial average of the double cropping pixels (annually) also showed a positive change in NIR under both RCP 4.5 (−2.35% in the 2020s to 1.98% in the 2080s over baseline NIR of 2.14 mm/day) and RCP 8.5 (−1.99% in the 2020s to 5.56% in the 2080s), suggesting an increase in the crop water requirement. It clearly showed a negative change during the Kharif season, while a relative rise in NIR was reported during the Rabi season. Among the models considered, MRI reported the least NIR (−6.53% relative change annually in 2080s under RCP8.5), while MPI reported the maximum change in NIR (annual increase in 13.09% in the 2080s of RCP 8.5) (Table 9). The spatial mean of the monthly NIR of the baseline period for the four models was compared with the IMD-estimated NIR for the same period (Table 10). The performance of the models and their ensemble was carried out using three performance indicators, namely, NSE, CRM, and R². The ensemble outperformed the individual models with NSE value of 0.96, CRM value of −0.04, and R² of 0.97, indicating a very close match. The models showing the best performance individually were found to be CNRM, followed by MPI, MRI, and CCSM, respectively.

An attempt was made to estimate NIR using a satellite-derived k_c layer for three cropping pixels, namely, Kharif-only, Rabi-only, and double/triple-cropping pixels over India. Using ‘ET₀ & NIR toolbar’, the HS estimated monthly ET₀ was generated for the present climate by employing IMD gridded temperature dataset. Due to the vast study area, collection of information on the specific type of crop grown and characteristics such as development stages, growth, water availability, etc., of each cropping area is impossible. Therefore, we estimated NIR using a satellite-derived k_c layer for three cropping pixels, namely, Kharif-only, Rabi-only, and double/triple-cropping pixels over India. Two satellite-based monthly k_{c_act} layers were prepared, namely, k_{c_EEFlux} and k_{c_FEWS}. The spatial and temporal output of NIR and its intermediate variables, i.e., ET_cm and R_e for the present and projected climate were generated using the prepared k_c layers. NASA's Earth Exchange Global Daily Downscaled Projections (NEX-GDDP) dataset was employed for projecting the future scenario. Four models, namely, CCSM, CNRM, MPI, and MRI were selected from the NEX-GDDP dataset, which includes downscaled daily projections for RCPs 4.5 and 8.5 to evaluate the future scenario.

The following conclusions could be drawn from the present study:

• 1.

  The HS method, although it slightly underestimated ET₀, could be employed in those areas where continuous meteorological data availability is limited (R² = 0.92). However, the reliability of the temperature dataset must be analysed to obtain accurate ET₀. Although there were discrepancies in the extreme values of FAO56-PM and HS estimated ET₀, both showed a similar trend in all months (R² = 0.88).

• 2.

  Both the prepared monthly k_c layers (k_{c_EEFlux} and k_{c_FEWS}) could adequately estimate ET_cm for the Kharif (Kharif-only pixels) and annual (double/triple pixels) estimations. However, k_{c_EEFlux} was found to highly overestimate the ET_cm for the Rabi season (Rabi-only pixels) even though it is Kharif and annual estimates were much closer to the previously published data compared to k_{c_FEWS}. Overall, we could say that the USGS-FEWSNET SSEBop model offered a much more reliable estimate in comparison with EEFlux for the present study area.

• 3.

  Assessing the NEX-GDDP projected precipitation and temperature, we observed a larger increase in precipitation during the Rabi season (28.76%) as compared to the Kharif season (22.85%) in the 2080s under RCP 8.5 with respect to the ensemble baseline. Our findings also showed a probable higher rise in temperature (maximum and minimum temperatures) during the Rabi season compared to the Kharif season under both RCPs 4.5 and 8.5.

• 4.

  Corresponding to the increasing temperature along the future time slices, an increasing estimate of ET₀ was generated under both RCP scenarios. Our results also indicated a greater increase in ET₀ during the Rabi season compared to the Kharif season.

• 5.

  The ET_cm, R_e, and NIR were estimated for three cropping pixels, namely, Kharif-only, Rabi-only, and double/triple under the projected climate change scenarios. The ensemble ET_cm indicated that the relative change with respect to the baseline for Rabi pixels was much higher than that of Kharif pixels under both RCPs 4.5 and 8.5. Correspondingly, the projected ensemble NIR indicated a negative relative change during the Kharif season, while a probable positive relative change was reported during the Rabi season.

• 6.

  Among the models considered, CNRM was found to be the most suitable model followed by MPI, MRI, and CCSM for the present study.

In addition, since the present study confined itself to a broad classification of three cropping pixels, estimations carried out (local or regional) for much finer land use would be a valuable contribution.

No funding was received for this study.

All the data used in this study are openly available in the public domain and the sources are indicated in Section 2.2.

ArcMap is used under license which needs to be procured from ESRI. The ArcET toolbar can be obtained from the authors free of charge by sending email request.

G.N. conceptualised the study, did data acquisition, prepared the methodology, executed the study, and wrote the original draft. A.B. conceptualized the study, edited the manuscript, and visualized the study. A.B. supervised the study and edited the manuscript.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.