1. Introduction
The development of groundwater irrigation in the past few decades has helped in securing crop yield in many semi-arid parts of the world, especially in India [
1]. This development has led to the over-exploitation and increasing depletion of the groundwater resources both in quantity [
2] and quality [
3,
4]. Remote sensing data for retrieving area under irrigation and its historical evolution are extremely useful for water management. However, use of optical satellite data in the case of Indian agriculture is challenging as (i) it is dominated by small farm holdings (average farm less than 1 ha), (ii) tropical regions with high cloud cover led to less availability of high-resolution satellite images with high temporal revisit during various cropping seasons, (iii) it has a large diversity of crops and agricultural practices that are difficult to capture by existing indices [
5,
6] and (iv) many of these crops have a short cycle of 3 to 4 months. Exploring new combinations of various optical indices with classification algorithms are promising way to improve the accuracy of irrigated cropland estimation.
Census-based statistical approach and remote sensing-based approaches are currently used to produce irrigated land statistics at National or regional scale, but results vary according to the data and methods used [
7,
8]. Broadly, these statistics reveal that groundwater irrigation in India has increased from 20 million ha to 60 million ha respectively from 1950 to 2000 and the net irrigated area has been increased from 22 million ha to 75 million ha for the same period [
9]. Uses of electricity and diesel for the groundwater pumping accounts approximately 16–25 million metric ton of carbon emissions, which is around 4–6% of India’s total carbon emission [
1]. Due to this uncontrolled groundwater pumping, groundwater levels are declining and getting disconnected from the surface, leading to the drying up of rivers observed in many places especially in the Deccan Plateau [
10,
11,
12].
However, monitoring of cropping pattern using satellite data is still a big challenge at the scale at which water resources have to be managed. i.e., village or small watershed level. Firstly, farm plots are very small. About 94% farmers in India are having a land holding of smaller or equal to 1 ha, which is continuously decreasing [
13,
14]. Monitoring these small-sized irrigated croplands is highly required to access irrigation water use [
15,
16]. Second, large temporal variations in plot size and practices are observed. Declining water table directly impacts on crop yield and land use practices [
7,
17,
18,
19]. Farmers develop various strategies to adapt to depleting water resource: either reducing the number of crop cycles in the same plot or reducing the irrigated area in the farm and use rainfed crops in the remaining part or even leave part of it as fallow land [
18]. Farmers are discouraged from growing more water consuming crops (e.g., rice, sugarcane, and banana) in arid and semi-arid zones due to environmental implications. Farmers can adopt the rainfed and seasonal crops (e.g., wheat, maize, sorghum, sunflower, and beetroot) which consume reasonably less water. Finally, under water scarce situations, farmers often abandon long duration crops (e.g., sugar cane, banana, turmeric) and grow short duration (3 to 4 months) cash crops with multiple cropping [
13]. All these factors increase the difficulty in assessing irrigated areas in the farming system.
Satellite image derived indices are commonly used for irrigated and non-irrigated cropland area identification. They are usually based on the detection of spatial differences in the presence of water and/or vegetation cover [
20,
21,
22]. Normalized Differential Moisture Index (NDMI) (also known as land surface water index (LSWI)or normalized difference water index (NDWI)) is sensitive to surface soil moisture [
20,
21,
23]. Liquid water has strong light absorption in the shortwave infrared (SWIR) band, which makes NDMI highly sensitive to the total vegetation water content [
24]. It is widely used for irrigated cropland classification [
25,
26]. It uses near-infrared (NIR, λ = 0.851–0.879 μm) and SWIR (λ = 1.566–1.651 μm) bands of Electro-Magnetic (EM) spectrum. For vegetation, the most widely used is NDVI (Normalized Differential Vegetation Index) which is sensitive to chlorophyll content [
27]. Normalized Differential Vegetation Index (NDVI) was generally used to identify the irrigated cropland and LULC (land use and land cover) classification at various scales using various spatial resolution satellite images [
22,
28,
29,
30]. Enhanced Vegetation Index (EVI) is similar to NDVI but has soil reflectance correction and canopy reflectance adjustment properties [
31]. EVI is highly susceptible to the vegetation and sometimes better than NDVI [
32]. The time series of NDVI and EVI are commonly used to represent seasonal rhythms and phenological variations for different land use types [
20]. The difficulty is that even in rainfed areas, spatial variations of moisture occur due to spatial heterogeneity of rainfall and soil properties. Similarly, in areas where a diversity of crops is present, spatial variations of vegetation cover might be due to timings of crop rotations, the difference in rooting depth for different species or presence of trees. For this reason, the use of a single index for detecting irrigation is increasingly questioned [
32,
33] for the case of humid regions. A multiple indices (vegetation, surface moisture, and surface temperature)-based approach is preferable for the irrigated and non-irrigated cropland classification [
34,
35].
Moderate Resolution Imaging Spectroradiometer’s (MODIS), Terra mission and Landsat satellite products are well known for land use and land cover studies [
6,
36,
37,
38,
39]. However, the available high temporal frequency satellite products (MODIS) are of coarse (250 m or more) to medium (Landsat (30 m)) spatial resolution. In a country like India where small (less than 1 ha) agricultural fields are dominant, Very High Spatial Resolution (VHSR) remotely sensed images are required to identify the irrigated crops [
5,
7,
18]. At the same time, extensive ground data is required for the calibration and validation of classification outputs. It is evident from the literature that several studies focused on identifying irrigated cropland areas [
7,
22,
37,
40]. For precise irrigated classification outputs either time series [
37,
41] for the complete cropping duration or in case of non-availability of time series satellite images, a single satellite image during the peak of cropping season [
40,
42] is required. A robust methodology capable of quantifying irrigated cropland with high accuracy at field scale from the satellite images remains a challenge.
Support vector machines (SVM) classifier is a supervised non-parametric statistical learning mechanism. It was developed in the late 1970s, but SVM became more famous for remotely sensed image processing in last decade due to algorithm’s flexibility and higher classification accuracy outputs [
37,
43,
44]. Support vector machine is a set of supervised learning methods widely used for the classification, regression and outlier detection [
45,
46]. Support vector machine was formerly known as binary classifiers with an algorithm adapted to reduce the multiclass problems [
47]. It is well known for the optimal separating hyperplane between various classes with the help of training cases. The training samples falling within the boundaries are the most critical for discrimination, SVM accurately illustrate these samples and prioritize them from the other classifiers. With comparison to other classifiers like neural networks, decision tree, and discriminant analysis, SVM performs better in the case of small training sets too [
47,
48]. The samples lie at the edge of the class distribution in feature space, SVM uses them as training samples, while maximizing the likelihood classifier uses class means and covariance of training samples as an input [
37,
47]. It fits for the remote sensing applications to perform precise land use and crop classifications [
49].
The objectives of this study were to assess the potential of optical indices derived from available cloud free multitemporal Landsat images for estimating the historical evolution of irrigated cropland area and to evaluate the impacts of this evolution on groundwater depletion in a semi-arid region of Southern India. For calibration, we used a large ground surveyed dataset of the Berambadi experimental watershed. The ground observations of irrigated and non-irrigated farm plots were done by surveying farmers’ fields from 1990 to 2016. It was noticed that the irrigated cropland classification requires time-series of vegetation indices [
37,
44,
50]. Moreover, it has been shown that classification accuracy is higher when using multiple vegetation indices vs. single vegetation index in case of using one single date image [
51]. One index was not able to consistently classify irrigated area using a single image, while in a combination of multiple indices high and consistent accuracy was observed. While applying this method to a watershed in India, we characterized the spatial heterogeneity in the dynamics of irrigation development across the watershed, and it links with the groundwater resource depletion.
4. Conclusions
Characterizing the spatiotemporal evolution of irrigated croplands is essential for sustainable resource exploitation in intensive agriculture areas. This study showed that available multitemporal remote sense data could be used to retrieve historical evolution of irrigation and its impact on groundwater resources. While we used Landsat multi-temporal imagery, it would be interesting in future studies to apply our approach on Sentinel 2 time-series to compare the results obtained with these two different types of datasets.
We have showed that the combined use of multispectral indices like NDVI (vegetation), EVI (soil impact correction to vegetation) and NDMI (moisture) could provide accurate classification of irrigated cropland. Using SVM algorithm (trained with intensive ground sample dataset) allowed us to generate irrigated cropland statistics for the past four decades in a semi-arid watershed in South India, which was a challenge considering with the small plot dimensions (less than 0.5 ha). The resulting irrigated and non-irrigated cropland classified maps were generated for rabi and summer seasons with high classification accuracy and kappa coefficient.
Results show the contrasted spatial distribution of irrigated cropland in the watershed and its evolution over time: while irrigation was developed very early in the valley areas in the East (downstream), it expanded recently to all the upstream areas while decreasing in the East. This was found consistent with the observed evolution of groundwater resources in the watershed: while in the East part groundwater was severely depleted since 2010, the decline in water table levels is more recent in the West.
The constructed irrigated cropland statistics can be used as an input to groundwater models for understanding the interactions between climate, agricultural practices, and water resources. They are also relevant to implement management strategies and support to policy making. Creating awareness among the farmers to adopt sustainable agriculture practices would perhaps the only way to escape from the onslaught of slow ecological disaster.