Warming temperatures exacerbate groundwater depletion rates in India

Abstract

Climate change will likely increase crop water demand, and farmers may adapt by applying more irrigation. Understanding the extent to which this is occurring is of particular importance in India, a global groundwater depletion hotspot, where increased withdrawals may further jeopardize groundwater resources. Using historical data on groundwater levels, climate, and crop water stress, we find that farmers have adapted to warming temperatures by intensifying groundwater withdrawals, substantially accelerating groundwater depletion rates in India. When considering increased withdrawals due to warming, we project that the rates of net groundwater loss for 2041–2080 could be three times current depletion rates, even after considering projected increases in precipitation and possible decreases in irrigation use as groundwater tables fall. These results reveal a previously unquantified cost of adapting to warming temperatures that will likely further threaten India’s food and water security over the coming decades.

Increased irrigation in response to warming temperatures has accelerated groundwater depletion rates in India

INTRODUCTION

Climate change and natural resource degradation are challenging agricultural production across the globe. This is particularly true in India, where climate change is projected to decrease the yield of staple crops including rice and wheat by up to 20% by midcentury (1, 2). Simultaneously, groundwater, which provides more than 60% of India’s irrigation, is being depleted at an alarming rate primarily because of water withdrawal for irrigation (3–6). Reduced water availability due to groundwater depletion and climate change (7, 8) is expected to reduce cropping intensity and production (9–11), which will challenge India’s food security and threaten the livelihoods of over one-third of India’s 1.4 billion people (12). This has implications for global food security given that India is home to 18% of the global population and is the second-largest producer of common cereal grains like rice and wheat (13).

To date, most studies on groundwater depletion and crop production in India have focused on the individual effects of climate change and groundwater depletion on crop production (9, 11, 14); however, the feedback mechanisms (Fig. 1) between these variables are not well understood. The majority of existing studies that aim to understand these feedback mechanisms use global climate models (GCMs) or land surface models, with known or unknown uncertainties arising from the use of complex modeling frameworks (15–17). These studies do not account for real-world farmer decision-making, including how farmers may adapt to changing climate through changes in irrigation decisions (6, 10, 18). These studies are also conducted at coarse scales because of a lack of high-resolution groundwater depletion (19, 20) and crop productivity data; coarse-scale studies (6, 17) may mask the important heterogeneity that occurs in countries like India, where irrigation practices, crops grown, and water use efficiency vary at fine scales (6, 21). Last, the majority of studies linking climate change with groundwater depletion in India have focused primarily on changes in precipitation patterns (4, 22) and how farmers respond to changes in rainfall (10). However, studies in different parts of the world have found that warming temperatures can increase groundwater depletion (23) by increasing crop water demand (24, 25), which results in farmers increasing irrigation use to meet this demand (26, 27).

Fig. 1. Conceptual framework showing the hypothesized feedback mechanisms between climate, groundwater levels, and crop water stress in India.

Crop water stress is expected to increase with an increase in potential evapotranspiration (PET) due to increased evaporative demand caused by warming temperatures. Farmers may adapt to this increased evaporative demand by increasing irrigation withdrawals, reducing groundwater levels (GWLs) and crop water stress. Increasing precipitation could increase GWLs by increasing groundwater recharge, while warming temperatures could increase evapotranspiration (ET) to reduce recharge and GWLs. We highlight the dependent variables considered in our regression analyses using the associated equation numbers from Materials and Methods.

We develop an empirical model that links groundwater depletion, crop water stress, and climate in India. We did this by developing a unique dataset that combines groundwater depth data from thousands of wells across India from 2004 to 2013 (figs. S1 and S2), a high-resolution (1 × 1 km²) remote-sensing data product that measures crop water stress (28) and weather data including precipitation and temperature. Using panel fixed-effect regressions (see Materials and Methods), we modeled the relationships between groundwater depletion, temperature, precipitation, and crop water stress during the two main agricultural growing seasons (monsoon and winter) in India (tables S1 and S2). In addition to modeling these relationships across all of India, we also tested whether the relationships differed for India’s two major aquifer systems (unconsolidated and consolidated aquifers; tables S1 and S2). Then, by combining our model results, projected rainfall and temperature estimates from 10 climate models (figs. S3 and S4 and table S3), and potential reductions in irrigation use as groundwater tables fall, we estimated potential changes in groundwater levels (GWLs) from 2041 to 2080.

RESULTS

Warming temperatures and falling groundwater tables increase crop water stress

We find that warming temperatures increase crop water stress in both the monsoon and winter growing seasons due to an increase in water demand, or potential evapotranspiration (PET). Increases in PET reflect increases in atmospheric water demand that occur due to warming temperatures, which in turn increase the amount of water lost from crops through evapotranspiration (ET). If farmers are able to compensate for this increased demand by increasing irrigation as an adaptation strategy, then ET would equal PET and the crop would experience no water stress. We find that 1°C warming is predicted to decrease ET relative to PET, our measure of crop water stress, by 5.9% in the monsoon season and by 3.5% in the winter season (Fig. 2A and table S1). The greater increase in crop water stress in the monsoon season is likely because the monsoon season experiences higher temperatures and radiative effects than the winter season, and this is associated with increased PET. In addition, we find that increased monsoon and winter precipitation reduces crop water stress (Fig. 2A and table S1), as increased precipitation is associated with an increase in recharge and available water supply for crops. Our results are similar across the two main aquifer types in India, though the effects of the monsoon season warming on crop water stress are larger for unconsolidated aquifer systems (table S1).

Fig. 2. Groundwater extraction rates increase in warmer seasons, helping to limit increases in crop water stress.

The figure shows associations between precipitation, temperature, groundwater depletion, and crop water stress from panel fixed-effect regressions. Panel (A) shows the effects of temperature, precipitation, and groundwater depletion on crop water stress and panel (B) shows the effects of temperature and precipitation on changes in groundwater depth. * and ** indicate P < 0.05 and P < 0.01, respectively. n ranges from 22,303 to 31,492 depending on the regression model. Error bars represent 95% confidence intervals derived as coefficient value ±1.96 × standard error of the regression coefficient (tables S2 and S3).

Considering the impacts of groundwater depletion on crop water stress, we find that falling groundwater tables are not associated with crop water stress in the monsoon season likely because the monsoon provides ample water for crops through precipitation and surface water availability. Groundwater depletion, however, is leading to increased crop water stress in the winter season, though this effect is small: For every 1-m decline in GWL, crop water stress increases by 0.3% (Fig. 2A and tables S1 and S4). The small effect of groundwater depletion on crop water stress suggests that farmers have not lost their ability to irrigate in response to short-term fluctuations in groundwater and have largely been able to meet increased water demand driven by increasing temperatures. This is particularly true in unconsolidated aquifer systems where the effect of groundwater depletion on crop water stress is even smaller (tables S1 and S4). These results are robust to potential selection bias (tables S5 to S7) and the choice of ET (table S8) products.

Warming temperatures exacerbate groundwater depletion

Given that our results suggest that farmers increase irrigation use to meet increased crop water demand as temperatures warm, we examined the impact of warming temperatures on groundwater withdrawal rates in India. We find that an increase in mean temperature by 1°C was associated with a reduction in groundwater gain in the monsoon season (25.11 cm) and an increase in groundwater withdrawal during the winter season (10.90 cm), even after controlling for the positive effects of rainfall on recharge (Fig. 2B and table S2). We find that the mechanisms explaining warming-induced groundwater depletion differ based on aquifer type; in consolidated aquifer systems, the warming-induced GWL declines are primarily driven by decreased monsoon recharge (29), whereas in unconsolidated aquifer systems, declines are largely due to increased groundwater withdrawal in the winter season (table S2). A 1°C warming in the monsoon and winter seasons could increase net groundwater depletion rates from 8.15 cm year⁻¹ (Fig. 3) to 36.01 cm year⁻¹ (table S2) across India. Considering a specific yield of 0.12 (3), this is equivalent to an increase in groundwater storage depletion rates from 0.98 to 4.32 cm year⁻¹. These depletion estimates are realistic given that they are within annual depletion rates reported for India (3, 4, 30–32) and are considered to be substantial given that they are on the higher end of depletion rates for other major groundwater depletion zones globally (33). These findings are consistent when observation-based GWL changes were replaced by those derived using terrestial water storage (TWS) data from Gravity Recovery and Climate Experiment (GRACE) satellites (see Materials and Methods and tables S9 and S10). Last, these results are robust to selection bias (tables S11 to S13) and the potential cooling effects of irrigation (table S14).

Fig. 3. Groundwater depletion is expected to increase under different future climate change scenarios.

The figure shows current and future net GWL change (cm year⁻¹) under different climate change scenarios. The net GWL change is the sum of GWL changes in the monsoon (+) and winter (−) seasons, which also accounts for potential groundwater savings that may occur if farmers reduce the amount of irrigation they use as groundwater tables fall. Parentheses along the x axis show the mean and SD of the difference in future minus current temperatures (°C) from 10 climate models across groundwater observation sites used in the regression models. Error bars for the current scenario represent 95% confidence intervals based on the standard errors of the mean values across the sample (n = 3192). Error bars for future climate scenarios represent ± one standard error of projected mean values from the 10 climate models.

Warming temperatures may triple future groundwater depletion rates

We estimated how much groundwater depletion may increase under future climate change scenarios that account for increases in temperature (24), increases in monsoon precipitation (34, 35), decreases in winter precipitation (36), and decreases in irrigation use that may occur as groundwater tables fall (11, 14, 24, 34, 35). Such an analysis is critical as sustainable management of groundwater resources requires accurate groundwater resource accounting and, to date, no studies have accounted for the potential increase in depletion caused by warming-induced increases in irrigation in India. Our results indicate a substantial decline in GWL in the future (Fig. 3 and figs. S4 to S7), as warming temperatures coupled with a decline in winter precipitation more than offset recharge from an increase in monsoon precipitation. Across climate change scenarios, we find that our estimate of GWL declines from 2041 to 2080 is 3.26 times current depletion rates on average [ranging from 1.62 to 4.45 times depending on the climate model and Representative Concentration Pathway (RCP) scenario; Fig. 3 and figs. S4 and S5]. Furthermore, our model suggests that climate change may expand where groundwater depletion occurs across the country, with increased depletion in northwest, southwest, and some parts of central India (Fig. 4 and figs. S2 and S8 to S12); this is of concern given that some of these regions are already facing high levels of groundwater depletion (3, 4). Overall, our results suggest that warming temperatures will substantially amplify groundwater depletion across India over the coming decades.

Fig. 4. Climate change will accelerate groundwater depletion across northwest, southern, and central India.

The figure compares the (A) 2004–2013 mean or current net GWL change (cm year⁻¹) with those predicted for (B) 2050 (i.e., 2041–2060 average climate) and (C) 2070 (i.e., 2061–2080 average climate) using regression results, and projected climate under the RCP2.6 scenario, and potential reductions in irrigation use that may occur as groundwater tables fall. The current GWL change map (A) shows an inverse distance weighted interpolated map of mean GWL loss for the 2004–2013 period across 1604 groundwater measurement locations.

DISCUSSION

We find that warming temperatures have accelerated groundwater depletion as farmers have increased the amount of irrigation used to meet growing crop water demand. We project under a business-as-usual scenario that continued warming may triple groundwater depletion rates over the coming decades. This is critical, given that more than 60% of the nation’s irrigated agriculture depends on groundwater, and portions of India are already facing severe groundwater depletion (3, 10, 37). While increasing irrigation use successfully minimizes the negative impacts of warming temperatures on crop water stress (17, 26, 27, 37), the resulting groundwater depletion could reduce farmers’ abilities to irrigate over decadal time scales. Our study reveals the previously unquantified cost of climate change adaptation that is crucial to consider when developing groundwater management strategies for India’s future food and water security.

Historically, farmers have been able to maintain groundwater irrigation as water tables fall largely because of policies that facilitated groundwater extraction and a largely ungoverned groundwater irrigation economy (7). Increased access to bore wells, free or subsidized electricity, and a lack of electricity metering have allowed farmers to withdraw groundwater on demand (38, 39), leading to overexploitation (3, 10). To reduce this overexploitation, effective policies are needed for rationing the power supply (40), metering electricity usage (38), regional water resource development and allocation (41), rewarding farmers that invest in groundwater recharge (40), and reducing or removing energy subsidies (42). In addition, groundwater-saving interventions, such as the use of efficient irrigation technologies (such as drip or sprinkler irrigation), cultivation of less water-intensive crops, and supplemental irrigation through canals (7) may also be needed. Previous work has shown that a portfolio of such policies can enhance groundwater conservation (43). While challenges remain in implementing new regulations and interventions across the hundreds of millions of households that face groundwater depletion (7), without such measures, our results suggest that groundwater depletion rates will likely accelerate under climate change.

Warming-induced groundwater pumping will also likely increase the area facing groundwater overexploitation in the future. Currently, most overexploitation of aquifers is concentrated in the northwest and south India (44), but our results suggest that overexploitation may expand to include aquifers in the southwest, the southern peninsula, and central India by 2050 (Fig. 4 and figs. S8 to S12). Such an expansion is of concern because south and central India have hard rock aquifers that are more difficult to recharge and have less storage capacity compared to the alluvial aquifers found in northwest India (5). It is therefore more likely that farmers in these systems will lose their ability to irrigate if aquifers become overexploited. Consequently, water-saving policies and interventions that are currently focused on northwest India need to consider south and central India. Targeting water-saving policies and interventions to these regions before substantial groundwater depletion occurs could help farmers maintain their ability to irrigate and cope with warming temperatures over the coming decades.

We used a panel fixed-effects linear regression to model historic and future associations between temperature, rainfall, and groundwater depletion. By using an empirical statistical model, we were able to account for real-world farmer decision-making and infer that farmers have increased the amount of groundwater irrigation used in response to warming temperatures. While we do not model groundwater price or groundwater access explicitly, by using empirical data, our model accounts for changes in groundwater price and access that have historically occurred as water tables decline. Our panel model also allowed us to better capture causal relationships between climate and groundwater depletion as site fixed effects removed the impact of any time-invariant factors that vary across wells.

We do acknowledge, however, that our projection is only a first-pass estimate and likely contains uncertainties arising from several factors. First, we model reductions in irrigation use using empirical values from previous studies that have used historical data to model changes in crop area, yield, intensity, and type that have occurred as groundwater tables fall (11, 14, 45). We do not account for potential changes that to date have not occurred at scale, such as the adoption of efficient irrigation technologies, the removal of electricity or minimum support price subsidies (46, 47), and the complete loss of groundwater irrigation in critically overexploited regions (9). Furthermore, we do not account for potential changes in recharge that may occur due to the water diversion through canals or changes in the future rainfall distribution. In addition, we do not account for potential increases in crop water use efficiency that may occur under elevated atmospheric carbon dioxide concentration (48). We also do not explicitly consider biophysical or environmental limits to groundwater withdrawals (49). Our work should therefore serve as an initial quantification of what the impacts of warming temperatures could be under business-as-usual conditions, which should be further refined in future work using process-based hydrological, agronomic, and socio-economic models that can account for such changes.

To conclude, our results reveal that farmers have adapted to warming temperatures by increasing groundwater irrigation use, which has led to a substantial acceleration of India’s groundwater depletion. Using our model estimates, we project that under a business-as-usual scenario, warming temperatures may triple groundwater depletion rates in the future and expand groundwater depletion hotspots to include south and central India. Without policies and interventions to conserve groundwater, we find that warming temperatures will likely amplify India’s already existing groundwater depletion problem, further challenging India’s food and water security in the face of climate change.

MATERIALS AND METHODS

Materials

Remotely sensed actual ET and potential ET data

Seasonal ET data were obtained from summing monthly ET derived using an ensemble surface energy balance model that has been validated in India (28). This product estimated ET as the ensemble mean from seven thermally driven surface energy balance models (28). Mean ET values were extracted and calculated using 3 × 3 pixels around each groundwater site that was classified as cropped agriculture [based on the MODIS (Moderate Resolution Imaging Spectroradiometer) Land Cover product (50) during the monsoon season or based on winter cropped area (51) during the winter season]. PET was estimated using the Penman-Monteith (PM) equation (52) based on meteorological inputs from the NASA MERRA-2 reanalysis product, the same dataset that was used to calculate ET (28) for consistency and to provide appropriate boundary conditions. We specifically used the standardized American Society of Civil Engineering (ASCE)-PM equation (53) for short grass as

$$\mathrm{PET}=\frac{0.408\text{Δ}(R_{\mathrm{n}}-G)+\mathrm{\gamma }\frac{C_{\mathrm{n}}}{Ta+273}u(e_{\mathrm{s}}-e_{\mathrm{a}})}{\text{Δ}+\mathrm{\gamma }(1+C_{\mathrm{d}}{u}_2)}$$ (1)

where R_n and G are net radiation and soil heat flux (both in MJ m⁻² day⁻¹), respectively. Ta is the mean daily air temperature (in degrees Celsius), u is the mean daily wind speed at a 2-m height (in meters per second), e_s and e_a (in kilopascal) are saturated and actual vapor pressures, respectively, Δ is the slope of the saturation vapor pressure–temperature curve, and γ (in kilopascal per degree Celsius) is the Psychrometric Constant, respectively, and C_n and C_d are coefficients for short grass (ASCE-EWRI, 2005). The daily PET value was aggregated to a monthly and seasonal scale, and the ET/PET ratio was calculated at the seasonal scale.

GWL data and aquifer type

GWL measurement data from November 2004 to May 2014 were obtained directly from the Central Ground Water Board (CGWB; Government of India, 2017) for more than 20,000 locations in India (fig. S1) across all three climate gradients (fig. S2) (54). We only considered groundwater measurement sites that were located within a village (55) where 100% of its irrigation came from groundwater were classified as a cropped agricultural pixel (defined above) and were located within consolidated or unconsolidated aquifer systems (CGWB, Ministry of Water Resources; www.indiawaterportal.org/data/groundwater-scenario-india) (fig. S1). We considered all wells that had at least two observations; we did this to maximize the number of wells used in our analysis since previous work has shown that only using wells with long time series available can lead to selection bias (31). This resulted in 1714 to 5132 sites for different regressions with a total of 3192 sites where both pre-monsoon and pre-winter GWLs were measured. These criteria were used to ensure that GWL changes across the sites were primarily driven by changes in groundwater irrigation (along with recharge from precipitation).

We defined consolidated aquifers as both consolidated and semi-consolidated aquifers in our study. GWLs were measured four times per year, and we only considered pre-winter (November) and post-winter/pre-monsoon (May) GWL to estimate changes in GWL during the monsoon and winter seasons (32), respectively, as these are the two main agricultural seasons in India. For each site, GWL values outside of 3.5 SDs of the site-specific temporal mean values were considered outliers and were not used in regression analysis. Note that this analysis only removed some extreme values for a specific site (<0.5% of data), maintained the spatial variation in GWL changes across India, and did not alter our key results.

Climate and control variable datasets

All climate data reported here pertain to seasonal (monsoon as June to September and winter as December to March) mean or sum values. Mean temperature (0.5° × 0.5°) and total precipitation (0.25° × 0.25°) were obtained from CRU (Climatic Research Unit) (56) and TRMM (Tropical Rainfall Measuring Mission) version 7 (TRMM_3B43) (57), respectively. Mean leaf area index (LAI) (1 × 1 km²) values were obtained from NASA’s MODIS LAI product (58). Mean LST (land surface temperature) data (1 × 1 km²) were obtained from MODIS Terra and Aqua day and night products (59, 60). For temperature and precipitation projections, we obtained downscaled monthly (~0.05°) simulations (61) for 10 GCM (table S3) that were included in the Coupled Model Intercomparison Project Phase 5 CMIP5 (62) project. We selected the 10 most widely used GCMs (table S3) for which the downscaled monthly temperature and precipitation data were readily available (61). We obtained downscaled monthly climate projections for three RCP scenarios (RCP2.6, RCP4.5, and RCP6.0) for the years 2050 and 2070, which represent 20-year average climate projections for the periods 2041–2060 and 2061–2080, respectively (the data are available at www.worldclim.org/data/v1.4/cmip5_5m.html). We then calculated the mean seasonal temperature and total seasonal precipitation for each GCM and RCP scenario. We did not use RCP8.5 as studies have suggested that this is no longer a plausible scenario for future climate change (63).

Global-gridded TWS estimates from GRACE satellite mission

GRACE datasets (64–66) provide monthly TWS anomalies relative to a 2004–2010 time-mean baseline and are produced by three centers: the University of Texas’s Center for Space Research, GeoForschungsZentrum Potsdam (GFZ), and NASA Jet Propulsion Laboratory. We used mean values of TWS changes from the three datasets. However, we confirmed that the results (i.e., regression coefficients) were similar when the individual products were used.

Methods

Regression analyses

Our main results are based on panel fixed-effect regression models that examine (i) the effects of GWL, temperature, and precipitation on crop water stress (Eqs. 2 and 3) and (ii) the effects of temperature and precipitation on GWL change (Eqs. 4 and 5). These regressions were run using data for all of India and subsets of the two main aquifer types across the country (unconsolidated and consolidated aquifer systems). To control for potential heteroscedasticity, robust standard errors were derived for all regression coefficients in all regressions. Full results from all regression models (key analysis and robustness checks) are presented in supplementary tables (tables S1, S2, and S4 to S14).

Crop water stress regressions

We ran monsoon and winter season regressions separately (Eqs. 2 and 3, respectively) to understand the effect of GWL, temperature, and precipitation on crop water stress

$${(100\times \frac{\mathrm{ET}}{\mathrm{PET}})}_{iy,\mathrm{monsoon}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1{\mathrm{GWL}}_{iy}+{\mathrm{\beta }}_{2}T{\mathrm{mean}}_{iy}+{\mathrm{\beta }}_3{P}_{\mathrm{\_}\mathrm{mon},iy}+\mathrm{year}+c_{\mathrm{i}}+{\mathrm{\epsilon }}_{iy}$$ (2)

$${(100\times \frac{\mathrm{ET}}{\mathrm{PET}})}_{iy,\mathrm{winter}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1{\mathrm{GWL}}_{iy}+{\mathrm{\beta }}_{2}T{\mathrm{mean}}_{iy}+{\mathrm{\beta }}_3{P}_{\mathrm{\_}\mathrm{mon},iy}+{\mathrm{\beta }}_4{P}_{\mathrm{\_}\mathrm{win},iy}+\mathrm{year}+c_{\mathrm{i}}+{\mathrm{\epsilon }}_{iy}$$ (3)

where 100 × $\frac{\mathrm{ET}}{\mathrm{PET}}$ is a measure of crop water stress (ranges from 0 to 100, where 0 = high crop water stress and 100 = full water demand met), where ET_iy is the seasonal ET (in millimeters) (28) and PET_iy is the PM-based potential ET (52) at i location and y year. GWL (in centimeters) is the GWL at the start of the season (May for the monsoon season and November for the winter season). P_{_mon} and P_{_win} (in centimeters) are the total precipitation in the winter and monsoon seasons, respectively, and Tmean (in degrees Celsius) is the mean seasonal temperature. “year” is the year time trend to capture the potential linear trend in groundwater depletion in India, and c_i is the site fixed effect. We included P_{_mon} in the winter regressions to control for the potential of enhanced soil moisture due to increased monsoon precipitation. Since Tmean is used to calculate PET and is therefore highly correlated with ET/PET, we conducted a robustness check where we removed Tmean from the regression; we find all remaining regression coefficients (table S4) to be similar to those from the original model, suggesting that the correlation between Tmean and PET is not driving our results.

GWL change regressions

We ran monsoon and winter season regressions separately (Eqs. 4 and 5, respectively) to understand the effect of temperature and precipitation on GWL change

$${\text{Δ}\mathrm{GWL}}_{iy,\mathrm{monsoon}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_{1}T{\mathrm{mean}}_{iy}+{\mathrm{\beta }}_2{P}_{\mathrm{\_}\mathrm{mon},iy}+{\mathrm{\beta }}_3{\mathrm{LAI}}_{iy}+\mathrm{year}+c_{\mathrm{i}}+{\mathrm{\epsilon }}_{iy}$$ (4)

$${\text{Δ}\mathrm{GWL}}_{iy,\mathrm{winter}}={\mathrm{\beta }}_0+{\mathrm{\beta }}_{1}T{\mathrm{mean}}_{iy}+{\mathrm{\beta }}_2{P}_{\mathrm{\_}\mathrm{mon},iy}+{\mathrm{\beta }}_3{\mathrm{LAI}}_{iy}+{\mathrm{\beta }}_4{P}_{\mathrm{\_}\mathrm{win},iy}+\mathrm{year}+c_{\mathrm{i}}+{\mathrm{\epsilon }}_{iy}$$ (5)

where ΔGWL_iy is the GWL difference (in centimeters) between the start and end of the winter season (i.e., GWL measurement on November of year_i − GWL measurement on May of year_i+1) and the monsoon season (i.e., GWL measurement on May of year_i − GWL measurement on November of year_i) at i location and y year. Tmean, P_{_mon}, year, and c_i are the same variables as defined for Eqs. (2 and 3. LAI is the seasonal mean leaf area index used to control crop production. P_{_mon} in the winter regression was used to control for the effect of changes in soil moisture and groundwater availability due to monsoon precipitation.

We summed GWL changes in the monsoon and the winter seasons to estimate the net annual change in GWL (centimeters per year) from May to May (i.e., pre-monsoon to the post-winter period of a given agricultural year). The term “GWL change” here refers to the change in preseason and postseason groundwater tables and hence is not directly comparable with groundwater storage depletion rates reported in studies using GRACE data (3, 4). We considered GWL change as the dependent variable in our regressions as it provides a more direct measure of irrigation withdrawal, which is our main hypothesized response to climate change (Fig. 1).

Future GWL change projections

Projections of GWL change in 2050 (average conditions of 2041–2060) and 2070 (average conditions of 2061–2080) were made on the basis of present and projected monthly temperature and precipitation data obtained from the climate change simulations of 10 GCMs (table S3) from CMIP5 (61, 67, 68). The differences between the projected and current climate from the 10 climate models are shown in fig. S3. To estimate future GWL change, the differences in temperature and precipitation from future and current scenarios were multiplied by the temperature and precipitation coefficients from Eqs. 3 and 4. The temperature and precipitation coefficients were only used if they were found to be significant (table S2). We then summed GWL change from the monsoon and winter seasons to estimate net annual GWL change. We ran separate GWL change projections using projected climate data from each of the 10 GCMs (fig. S4), and then calculated the ensemble mean across all outputs (i.e., all projected GWL changes; figs. S4, S5 to S7, S9, and S10).

To account for potential decreases in irrigation use that will likely occur as groundwater tables fall, we used estimates found in other empirical studies that examined the historical relationship between groundwater depletion, crop production, and water use efficiency. These studies have shown that as groundwater tables fall in India, crop production and cropping intensity decrease and farmers switch from rice to less water-intensive crops (11, 14, 45). First, we incorporated maximum production loss estimates from studies which indicate that farmers in India have experienced up to a 2% loss in agricultural production (area × yield) during the monsoon season and up to an 8% loss during the winter season for every 1-m decline in groundwater (11, 14). Furthermore, we accounted for potential crop switching using values from a previous study (11) that found that farmers decrease the proportion of area under water-intensive crops by 0.2% for every 1-m decline in groundwater. We converted this to water-savings by assuming that this switch would lead to a 33% reduction in irrigation use as found in previous research (69). To incorporate the impact of these potential decreases in irrigation use into our scenario analysis, we added the amount of GWL change spared from these changes to our modeled GWL change estimates (Figs. 3 and 4).

A simple threshold method of negative GWL changes in the future was used to identify areas facing future depletion at both the point scale (n = 3192 GWL observation sites) (fig. S1) and the raster scale (created using MODIS land cover maps, interpolated inverse distance weighting maps of current depletion rates, and gridded climate projections; Fig. 4 and figs. S9 to S12).

Robustness checks

We ran several additional analyses as robustness checks. We find that our results are robust to all checks, meaning that the analyses described below produce similar values as our main regression results.

First, to check for selection bias, we reran all key regression models (Eqs. 2 to 5) using a different definition of groundwater irrigated sites. Specifically, we defined groundwater irrigated sites as those where groundwater measurement sites were located within a village (55) that had more than 50 to more than 90% (using increments of 10%) of its irrigation coming from groundwater (instead of 100%, which was used in the main analyses) (fig. S13 and tables S5, S6, S11, and S12).

Second, we reran our main regression models across only deep wells (average preseason GWL > 8 m) (14) to test whether our results (tables S7 and S13) are similar when we remove the possibility of including well depth data from nonirrigation wells located in shallow or perched aquifers (i.e., aquifer above the regional water table) (31).

Third, to ensure that our selection of ET products did not influence our results, we reran the crop water stress regressions using ET and PET products from the Global Land Data Assimilation System (GLDAS) (70) and MODIS (table S8) (71).

Fourth, we created a new panel dataset using TWS change data (65) from GRACE and aggregated all other variables (ET, PET, temperature, precipitation, and LAI) at the level of a GRACE pixel (~100 km). To simulate monsoon and winter season fluctuation in GWL, we averaged monthly changes in TWS and divided them by a specific yield of 0.12 (3). The use of a static value of specific yield is a widely used approach (3, 72, 73) because spatially explicit data on specific yield (74) are limited even in data-rich regions such as the U.S. aquifers (72, 75). We reran our GWL change regression models (Eqs. 4 and 5) using this new dataset and compared the regression coefficients to our original results. Results that use TWS change data from GRACE corroborate our main findings and suggest that warming temperatures will lead to a significant decline in groundwater depletion. For example, from this regression, we find that a 1°C increase in temperature will lead to groundwater depletion of about 6.6 cm in the winter and 20 cm in the monsoon season, and these effects are significant (P < 0.001; table S9). The results are also consistent whether we select GRACE pixels overlaying the groundwater observation sites (table S9) or all GRACE pixels across India (table S10).

Fifth, we conducted several robustness checks to assess whether changes in GWL-influenced temperature are due to the cooling effect of irrigation (76–78). Specifically, we ran a regression (Eq. 6) to test how LST (mean of daytime and nighttime LST) and air temperature are influenced by irrigation, as measured using the difference in GWL (post-winter minus pre-winter). The regression model is designed as

$${(T_{\mathrm{s}}-T_{\mathrm{a}})}_{iy}={\mathrm{\beta }}_0+{\mathrm{\beta }}_1{({\mathrm{GWL}}_{\mathrm{post}-\mathrm{pre}})}_{iy}+\mathrm{year}+c_{\mathrm{i}}+{\mathrm{\epsilon }}_{iy}$$ (6)

where T_s − T_a is the near-surface temperature difference, T_s is the seasonal mean daily LST (in degrees Celsius) estimated as the mean of daytime and nighttime LST from MODIS Terra or Aqua sensor, T_a (in degrees Celsius) is the seasonal mean air temperature from CRU, GWL_post−pre is the difference in GWL from postseason minus preseason, i and y are the site location and the year, “year” is the linear time trend, c_i is the site fixed effect, and ԑ_iy is the error term. We find that increased water withdrawal (greater change in GWL) is associated with a reduction in LST; however, the magnitude of change is small or not significant (table S14). We, therefore, do not believe that the effect of cooling from irrigation is strongly affecting our regression results. As an additional check, we assumed a 0.5°C cooling based on values found in other highly irrigated systems (77, 78) and estimated projected GWL changes, and find that future GWL will still decrease (fig. S14).

Supplementary Materials

This PDF file includes:

Figs. S1 to S14

Tables S1 to S14