Probabilistic flood risk mapping for Indian megacities using a scalable machine learning framework in Google Earth engine

Intensifying climate change and rapid urbanization pose a severe flood risk to India's major cities. Conventional flood assessment methods, however, are often geographically limited and fail to quantify critical uncertainties, hindering effective disaster management. To address this gap, we introduce the ALERT (Advanced Learning for Enhanced Flood Risk Prediction) framework, a scalable and integrated solution for flood analysis across the chosen Indian mega cities implemented entirely within Google Earth Engine (GEE). The framework leverages multi-source remote sensing data—including Sentinel-1 SAR and Sentinel-2 optical imagery—to perform robust Land Use/Land Cover (LULC) classification using a Random Forest algorithm. Its core innovation is a Bayesian probabilistic model that integrates SAR-derived floodwater evidence with topographic wetness index (TWI) and precipitation (ERA5) data to generate per-pixel flood probability maps, explicitly quantifying prediction uncertainty. Across five diverse megacities (Chennai, Mumbai, Bengaluru, Kolkata, Gurugram), the Random Forest classifier achieved LULC accuracies exceeding ∼90%. Critically, integrating this LULC context reduced flood detection false positives by ∼10-15% compared to traditional SAR-only methods. The Bayesian probabilistic approach further improved flood mapping accuracy by ∼3-5% over deterministic thresholding, providing a more reliable assessment of inundation risk. The validated ALERT framework delivers a replicable and scalable solution for dynamic flood monitoring, offering critical insights for risk-informed urban planning and enhancing climate resilience in vulnerable urban regions globally, with demonstrated applicability across strategically selected Indian megacities. • Random Forest identified as optimal for LULC classification in five diverse Indian cities (∼90% overall accuracy). • LULC integration quantitatively shown to reduce flood detection false positives by approximately 10-15%. • Novel Bayesian probabilistic flood mapping in GEE improves accuracy by ∼3-5% over deterministic methods. • Framework explicitly quantifies uncertainty and assesses flood risk under a +10% precipitation scenario. • Delivers scalable tools for operational flood monitoring and risk-informed urban planning.