Stacked ensemble model for flood layer extraction using EOS-04 satellite in Fine Resolution Stripmap (FRS) mode

This study presents a novel methodology for enhancing flood detection in India using Fine Resolution Stripmap (FRS) mode imagery from the EOS-04 Synthetic Aperture Radar (SAR) satellite. The proposed approach integrates deep learning and classical machine learning techniques through a stacked ensemble framework designed for water body extraction. Five deep learning models are independently trained, and their outputs are combined to construct a new feature matrix, which is subsequently used as input for five traditional classifiers: Logistic Regression, XGBoost, Random Forest, K-Nearest Neighbors, and Support Vector Machine. Among these, the XGBoost classifier achieved the highest classification accuracy of 96.8 %. A Land-Use-Land-Cover (LULC) water mask, provided by the National Remote Sensing Centre (NRSC), was applied to delineate the flood layer. Model performance was validated through cross-comparison with flood maps derived from concurrent-date optical satellite imagery, yielding an overall accuracy of 96.5%. Furthermore, a pixel-wise uncertainty map was derived from the ensemble predictions to quantify model confidence, offering additional insight into spatial reliability of flood extent mapping. By leveraging the complementary strengths of deep and traditional learning approaches, this method addresses the inherent limitations of FRS-mode SAR data for flood detection. The proposed framework offers a robust and scalable solution for operational flood monitoring, with significant implications for disaster preparedness and response across flood-prone regions of India.