How Satellites and AI Reveal Drought in India?

Droughts strain agriculture, water supplies and ecosystems. To act fast, researchers fused Sentinel‑2 multispectral images with machine‑learning algorithms. They examined 12 spectral indices – from leaf colour to soil moisture – across three Indian regions (Jodhpur, Amravati, Thanjavur) during the Rabi season (October‑April). Reference data came from official drought declarations and news reports, building a district‑by‑year “drought / no drought” table (2016‑2025). This ground truth labeled each satellite observation with 1 (drought) or 0 (no drought). Tested models – Random Forest, XGBoost, Bagging, Gradient Boosting – reached up to 84.8 % accuracy; a seasonal majority‑vote raised precision to 96.7 % for XGBoost and Bagging. SHAP analysis highlighted the Normalized Multi‑band Drought Index (NMDI) and Day of Season as the top predictors, followed by RECI, NDWI, NDMI and RDI. An ablation study showed that using the top five features maximizes performance, especially for XGBoost. These findings prove that satellite data combined with smart models can act as an early‑warning system for droughts, helping farmers and authorities plan better responses.

The study leverages Sentinel‑2, an ESA satellite constellation delivering 10‑30 m resolution imagery every five days. Analyzed spectral indices include NMDI, which blends red, near‑infrared, and short‑wave infrared bands to estimate the fraction of dry soil, and NDWI, sensitive to vegetation water content. Day of Season (DOS) is a temporal variable capturing phenological progression, crucial for distinguishing water stress from normal seasonal changes. XGBoost, a gradient‑boosting tree algorithm, optimizes a loss function via gradients while handling non‑linear feature interactions. Its built-in regularization (L1/L2) curbs overfitting, explaining its superior performance over Random Forest. SHAP analysis provides local importance values, allowing interpretation of each index's impact on individual predictions. The seasonal majority‑vote strategy aggregates monthly forecasts, reducing temporal noise and enhancing robustness. Comparing models with top‑k feature selection shows that even with the first two indices (NMDI and DOS), metrics exceed 80 %, but adding RECI, NDWI, NDMI, and RDI pushes precision to its peak, demonstrating the complementarity of vegetation, soil moisture, and chlorophyll indicators. These findings suggest that an integrated pipeline—satellite data, ground truth from official sources, and interpretable ML models—can be scaled to other vulnerable regions, offering timely decision support for water resource management.