Cropland classification and water stress vulnerability assessment in arid environment of Churu district, India using machine learning approach

Abstract

The focus was to map the cropland area and assess the water stress vulnerability within the area. Cropland mapping was performed via a machine learning (ML)-based ensemble classifier (EC). The leveraging of Random Forest, Extreme Gradient Boosting, and Support Vector Machines (RF, XGB, and SVM) models using Landsat 8 data of period 2022–23. The key inputs for the models included the spectral indices such as Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), Modified Soil Adjusted Vegetation Index (MSAVI), and Land Surface Temperature (LST) in the pre- and post-harvest period of both rabi and kharif crops. Kendall tau-b trend analysis (1991–2023) of the same indices was performed to estimate the long-term changes. The water stress was modeled via the generalized additive model (GAM). The EC identified 11,600.01 km² of cropland and 2254.32 km² of non-cropland, with a more than 90 % F1 score, 92.5% overall accuracy, and a Kappa coefficient (0.84). The trends show significant positive change for NDVI, EVI, and NDWI, while LST increased. The GAM demonstrated a strong fit, with an adjusted coefficient of determination (R²) of 0.89. Model diagnostics show an R² (0.79). The five-fold cross-validation confirmed the model's robustness. Moran's I analysis reveal a significant spatial clustering. The study concludes that water stress is influenced by spatially correlated factors, providing a framework for targeted crop management efforts in the area of Rajasthan, India.