Flood Detection from Sentinel-1 SAR Image using Multi-Scale Features and Distance-based Classification
Floods are some of the most devastating natural disasters that damagingly affect infrastructures, agriculture, and human life. The traditional method of flood mapping is mostly based on optical satellite imagery, which is affected by severe drawbacks such as blockage by clouds, inadequate lighting and weather conditions during heavy rainfalls conditions when such events are most direly needed, flood assessment-wise. The paper proposes an automated flood monitoring system based on Sentinel-1 Synthetic Aperture Radar (SAR) imagery in urban flood settings, which is known to have known limitations in the complex backscatter geometry. The proposed methodology presents Multi-Scale Flood Feature (MSFF) extraction, a technique that calculates the mean backscatter intensity and local variance statistics in 3×3, 5×5 and 7×7 adaptive neighbourhood windows, providing a feature descriptor of six dimensions per pixel. A distance-based prototype classifier classifies flood or non-flood by the minimum Euclidean distance to class-specific mean feature prototype with a normalised confidence score based on the prototype distance margin. The framework has two modes: dataset-level analysis to quantitatively evaluate performance and single-image inference to map floods in real-time. The experimental results indicate a total classification accuracy of about 70 percent and steady prediction of the probability of flooding across the levels of severity.