With one of the largest populations in the world, mainly composed of rural inhabitants who depend on agriculture as their principal income source, the competition for land and water throughout India is huge and expanding. Indian agriculture is a striking example of the long-term land use changes that have taken place in South Asia during recent decades: irrigation practices have been in general use for more than 50 years, supported by the Green Revolution [1
]. Yield improvements due to the increase in irrigated areas have been sustained by the construction of surface water tanks and the increase in the number of drilled irrigation wells from 0.15 to 20 million between 1960 and 2000 [2
]. Currently, Indian agriculture consumes more than 80% of the water used in the country [3
]—for instance, 89% in 2000 [4
], of which 65% is provided by groundwater [5
]. This estimate fell to 39% in 2000 according to [4
]. Indeed, these amounts vary widely from year to year. Investigations based on the Gravity Recovery and Climate Experiment mission (GRACE satellite) data estimated a decline of the groundwater stock in the Ganges basin in north India of 4.0 ± 1.0 cm yr−1
equivalent height of water (17.7 ± 4.5 km3
], 54 ± 9 km3
] for the 2000s, or more recently only 2 cm yr−1
during the 2002 to 2015 period [8
]. On the other hand, no clear trend appears in South India [9
]. These two contrasted results are mainly driven by the size of the groundwater resources: the presence in the north of the highest-yielding deep aquifers known in the world [11
], contrasts with the southern aquifers, which consist of shallow to moderately shallow fractured hard rocks with low porosity and storage capacities [12
]. Whereas in the north water extraction for irrigation purposes leads to a continuous decline in groundwater stocks, South Indian farmers experience recurring shortages, as shallow aquifers are temporarily emptied during periods of overexploitation when extraction is higher than recharge, and are partially refilled after heavy monsoon rainfalls. In this context, no overall declining trends were recorded by the GRACE mission [10
Apart from these long-term agricultural land cover and water use trends, land cover and water resources are highly correlated and variable particularly within a short-term range. Indeed, Indian monsoon precipitations are highly variable in both space and time, with magnitudes apparently linked to El Niño events [13
], their variations in the North being covariant with Indian ocean warming [8
]. These variations have a direct impact on the fluctuations of surface water through river-drained runoff captured in surface reservoirs equipped with large and small dams, but also cause groundwater fluctuations through aquifer recharge from the infiltration of non-evaporated water. Groundwater fluctuations are also strongly impacted locally by water extraction for agriculture. The best water extraction proxy is the extent of seasonal irrigated areas located around wells [15
In many parts of South India, the groundwater contribution to irrigated water has risen by 90%, encouraged by subsidized electricity for farmers. This financial support changed the status of this water resource from an emergency supply, in case of surface water scarcity, to the main water source for irrigated crops [16
]. To cope with the recurrent groundwater shortages associated with deficient monsoon rainfalls and overexploitation, farmers modulate their consumption by limiting the extent of irrigated areas. In Telangana State, Hyderabad region, the areas of inundated rice fields vary widely with the crop seasons (government statistics and National Remote Sensing Center estimates published on the Bhuvan web portal). Eighty percent of groundwater abstraction is dedicated to rice [15
]. The growing period is short (three months) and the rice areas represent a small proportion of the total. For instance, they have been evaluated at around 10% and 7%, respectively, for the monsoon season (Kharif) 2009 and the dry season (Rabi) 2010 [17
The quantification of irrigated and inundated areas from optical high resolution spatial remote sensing (SPOT, from French spatial agency CNES; ResourceSat from Indian Space Agency ISRO) in 2009–2010 provided spatial estimates of seasonal groundwater extraction for irrigation, using the daily irrigation practices associated with each seasonal crop type [15
]. The agro-hydrological model SWAT (Soil Water Assessment Tool [20
]) was adapted to simulate this daily extraction over several years, together with the aquifer recharge and runoff derived from the climate forcing variables [15
]. This methodology, based on an agricultural land use assumed to be fixed, derived from a set of two seasonal land covers (2009–2010) and allowed the simulation of groundwater fluctuations over a decade. It spatially assessed the groundwater availability and quantified theoretical groundwater shortages, i.e., when the crop water demand based on the satellite monitoring exceeded the groundwater storage [15
]. These studies quantified the contribution to aquifer recharge of the small surface reservoirs spread along the non-perennial river network that constitutes a rainwater harvesting system maintained by farmers. The main limitation of these studies is that they rely on fixed land use and thus do not represent the short-term modulations of irrigated crop areas implemented by farmers to cope with water shortages.
The recently launched Sentinel satellites are especially of interest for their acquisition strategies. Sentinel-1 (satellite S1A, launched in 2014; and S1B in 2016) and Sentinel-2 (S2A launched in 2015 and S2B in 2017) are the first generation of operational satellite EO missions for both optical multi-spectral and radar C-band detection of continental surfaces at a global scale, with high spatial and temporal resolutions (10 to 20 m; five to 10 days revisit interval), under a free access license.
In India, high-resolution satellite missions (56 m spatial resolution) have been used to measure seasonal land cover. In 2016, the National Information System for Climate and Environment Studies (NICES) project released a national database of seasonal net sown areas (http://bhuvan.nrsc.gov.in/data/download/index.php
), which highlights the high spatiotemporal heterogeneity of land cover and its temporal dynamics. This dataset is published as a fraction of the crop sown at 5 km resolution by the National Remote Sensing Center (NRSC) laboratory, part of the Indian Space Research Organization (ISRO). A similar program, called Water Bodies Fraction (WBF) and based on the same data type, produces the areal fraction of surface water every 15 days since 2012. The short-term variations of sown areas show a high temporal variation over recent decades in the Hyderabad region, with 41 to 58% of sown areas in Kharif and 8 to 22% in Rabi between 2005 and 2016. The seasonal restitution of the areas of inundated rice and other irrigated crops from multi-temporal high resolution satellite remote sensing is thus a promising method for estimating the surface water and groundwater requirements for irrigation.
Machine learning approaches are a way to produce land cover maps from remote sensing time series. More specifically, supervised classification methods, which require a training dataset, are suitable for exploring these massive amounts of data [21
]. Less automated strategies based on user-defined thresholds to provide more accurate indices can also be envisaged, but they are more time-consuming. In this agro-hydrological context, a threshold algorithm used to classify inundated rice shows higher accuracy than supervised random forest classification [22
]. However, the addition of radar images improves this automatic classification tree method. This last method has been found to be fast and to explore automatically and randomly the different spaces of features, here bands of satellite images time series [21
The aim of this study is to evaluate methodologies, based on new observations from the Sentinel-1 and Sentinel-2 satellites, for monitoring the essential agro-hydrological variables of this setting. These essential variables are (1) the seasonal areas of inundated rice, (2) the areas of irrigated crops, and (3) the dynamics of surface water areas within the water harvesting system. These three variables involve only a few percent of the total area and vary rapidly. This study considers two spatial resolutions, 10 and 20 m, to identify the value of preserving spatial rather than spectral resolution in this agro-system context. It explores the benefit of the synergy between Sentinel-1’s radar backscatter advantages (not sensitive to the persistent cloud cover during the Kharif season, highly impacted by crop growth and surface water) and the Sentinel-2’s multi-spectral detection, both at an appropriate spatial resolution (10 to 20 m). The estimates of surface water area will be compared to existing datasets from ISRO-NRSC, to discuss their interest for better understanding the contribution to aquifer recharge of the thousands of small surface reservoirs that constitute the water harvesting system. We have chosen the Random Forest algorithm (RF) provided with the Orfeo ToolBox (OTB) because it is a fast, open-source processor of high-resolution optical, multispectral and radar images at the terabyte scale. Highly appropriate for building a processing chain, it seems suitable for automatically producing agro-hydrological variables at a large scale [23
]. The seasonal land covers produced in this study have been used to estimate the seasonal Irrigated Water Demand (IWD), based on irrigation practices observed in previous studies [15
Flooded rice areas and surface water dynamics are essential variables in predicting groundwater usage in a small South Indian watershed. The first Sentinel-1 and -2 time series are used to estimate the accuracy of retrieval of irrigated and flooded sown crop areas for three contrasted season: the dry Rabi 2016 (R2016, December 2015 to March 2016), the monsoon season, known as Kharif 2016 (K2016, July to November 2016) and the next Rabi season in 2017 (R2017, December 2016 to March 2017). The spatial resolution at 10 or 20 m and the types of sensors are combined into six datasets to explore the retrieval accuracy of each variable.
Small irrigated areas are detected during the dry conditions of R2016 and K2016, as 3.5% and 5%, respectively, of the irrigated areas, with moderate classification confusion. A large increase of irrigated areas (16% of rice and 6% of irrigated crops) is estimated with low classification confusion in R2017. This is explained by the replenishment of water resources during the 2016 monsoon. This high seasonal variability represents the direct impact of the farmers’ strategy of adaptation to water storage.
In Rabi 2017, Sentinel-1 only, Sentinel-2 only and Sentinel-1 and 2 datasets show good classification accuracy (F-score for rice above 0.90) although Sentinel-1 used alone leads to a 10% overestimation of the rice areas (lower precision at 10-m resolution). In contrast the Kharif cotton areas seem to be accurately detected by Sentinel-1 only; their estimated extent remains similar (around 20%) after introducing Sentinel-2 (only four cloud-free images during the Kharif). The value of 20% appears highly realistic, as a similar area was found for Kharif 2009 [17
Sentinel-1 and 2 in Kharif and both Sentinel-1 and 2 and Sentinel-2 alone in Rabi seem to give the best estimates. We combined the results from these two datasets at both 10 and 20 m resolution to compute the water demand uncertainty for each season: 49.5 ± 0.78 mm (1.5% uncertainty) in Rabi 2016, 44.9 ± 2.9 mm (6.5% uncertainty) in the Kharif season and 226.2 ± 5.8 mm (2.5% uncertainty) in Rabi 2017. These amounts should be compared to seasonal aquifer recharge variation over the past decade (10 to 160 mm/year). The confusion over detection of irrigated crops in small areas in Rabi 2016 leads to a small uncertainty in water requirement estimates.
The 20-m resolution should be selected when scaling this methodology up to larger areas. The synergistic use of S1 and S2 increases the confidence level regarding the retrieval of irrigated areas.
The automatic detection of surface water dynamics is highly accurate with both Sentinel-1 and Sentinel-2 cloud-free images. Sentinel-1 is preferred, since it does not require any filtering step in the presence of clouds. The sizes of each surface reservoir are small within the study site, and account for 3% of the total area: the high resolution of Sentinel-1 allows for the detection of smaller areas; the surface water fraction found is then bigger than that estimated by the lower resolution of ResourceSat. Furthermore, surface water dynamics were monitored at the revisit periodicity of Sentinel-1 during July and August, which is not possible with optical remote sensing. This essential variable is important for calibrating runoff in spatially distributed agro-hydrological models, as well as for estimating the amount of infiltration from the bottom of the rainwater harvesting system back into the aquifer.