Floods remain one of the most destructive hydro-meteorological hazards, particularly within monsoon-dominated regions where rapid rainfall accumulation, river overflow, and changing land-use patterns collectively intensify flood vulnerability. Kolhapur district in Maharashtra, India, experiences recurring flood events driven by extreme monsoon precipitation, complex river networks, and low-lying floodplains. Conventional flood assessment techniques relying on optical satellite imagery frequently encounter limitations during flood events because persistent cloud cover and adverse weather conditions reduce image availability precisely when rapid inundation assessment is most critical. Synthetic Aperture Radar (SAR) remote sensing has emerged as an effective alternative because microwave signals penetrate cloud cover and operate independently of daylight conditions, enabling continuous flood monitoring throughout severe weather events. Simultaneously, advances in machine learning have significantly enhanced the capability to extract complex spatial relationships from multi-source geospatial datasets, improving the accuracy and efficiency of flood inundation mapping.

This study proposes a machine learning–driven framework for multi-temporal flood inundation mapping and spatial analysis in Kolhapur using SAR remote sensing observations. The framework integrates Sentinel-based SAR observations, meteorological information, terrain characteristics, groundwater conditions, and multi-temporal environmental variables to establish a comprehensive flood prediction workflow. Random Forest serves as the primary classification algorithm because of its robustness against noisy datasets and capability to model nonlinear relationships, while TensorFlow-based computational infrastructure supports scalable model development and optimization. Multi-temporal SAR observations enable the identification of persistent, seasonal, and event-specific inundation dynamics, allowing improved discrimination between permanent water bodies and temporary floodwater.

The proposed methodology consists of systematic data preprocessing, SAR image calibration, terrain correction, feature extraction, machine learning training, classification, temporal validation, and spatial interpretation. The framework further incorporates hydrological and climatic information to improve predictive reliability under varying monsoon conditions. Spatial analysis is employed to evaluate flood persistence, frequency, and geographical distribution across multiple flood seasons. The analytical workflow demonstrates how machine learning can substantially improve flood mapping accuracy while reducing dependence on manual interpretation. Similar machine learning-driven flood susceptibility modelling has shown promising performance in data-scarce environments, providing important methodological guidance for this study (Ganjirad and Delavar, 2023).

The research contributes a comprehensive conceptual framework that combines remote sensing observations with intelligent data-driven modelling for regional flood assessment. Beyond improving flood extent mapping, the proposed approach provides valuable support for disaster preparedness, emergency response planning, infrastructure protection, and sustainable watershed management. The study further highlights the importance of integrating artificial intelligence with geospatial technologies to strengthen climate resilience in flood-prone regions.