Monitoring Heat Extremes over India Using Earth Observations and Data Driven Approaches

Remote sensing enables spatially continuous and timely monitoring of hydro-climatological extremes by capturing key land–atmosphere variables across large regions, including for data-scarce areas. The rising frequency of heat extremes across India in recent decades underscores the need for effective monitoring, especially in data-scarce regions. This study evaluates the potential of monitoring heat extremes over the Indian sub-continent using satellite based observations and data driven approaches. For this, MODIS land surface temperature (LST) along with NDVI, land use/land cover and elevation information is used with traditional machine learning models namely Random Forest (RF) and XGBoost. Subsequently, the performance of the two ML models in estimating maximum temperatures across the Indian subcontinent was evaluated and validated using in situ temperature observations from the Indian Meteorological Department. Heat extremes were identified using both absolute temperature percentile thresholds and Standardized Temperature Index based heat stress categories. The performance of ML models was evaluated using station‑wise categorical verification metrics such as hit rate, false alarm ratio, and critical success index. Results show that the ML models exhibit higher accuracy in predicting mean temperatures compared to extremes, and XGBoost outperforms the RF model with lower RMSE and higher R². The results further reveals that ML model prediction skill exhibits considerable geographic variability across the sub-continent, with reduced performance over mountainous areas. This study demonstrates that integrating satellite-based data with machine learning provides an effective approach for monitoring heat extremes across the Indian subcontinent, particularly in data-scarce environments.