Abstract

Flood forecasting in monsoon-dominated river basins is challenged by sparse hydrological observations, rapid urban expansion, and strong seasonal variability. This study presents a multi-year machine-learning framework (2017-2025) that integrates Sentinel-1 SAR flood extents, CHIRPS and IMD rainfall, ERA5-Land hydrometeorological variables, and SRTM-derived physiographic attributes to predict flood inundation in Kolhapur, India. Random Forest (RF), Support Vector Machine (SVM), and Multi-Layer Perceptron (MLP) models were trained using progressively expanding temporal windows to evaluate the effect of multi-year hydroclimatic diversity on predictive skill. RF consistently achieved the highest accuracy and spatial coherence, with strong generalization to the independently withheld 2025 flood event. The multi-year approach captures monsoon variability-including extreme floods, prolonged wet spells, and deficit rainfall years-allowing the models to learn interannual rainfall-runoff-inundation relationships. Antecedent hydrological indicators such as cumulative rainfall, runoff, and soil moisture were identified as key predictors enabling short-range flood forecasting before inundation occurs. Beyond the local case study, the workflow demonstrates strong transferability to other monsoon-affected and data-scarce basins because it relies on globally available satellite datasets, terrain-based predictors, and non-parametric ML algorithms. The methodology is operationally scalable and adaptable for flood early-warning systems in rapidly urbanizing river basins worldwide.

Keywords: Flood inundation mapping, Machine learning, Synthetic Aperture Radar (SAR), Hydrometeorological modelling, Urban flood forecasting