Digital elevation model and SAR imagery — A marriage for effectual flood detection using novel lightweight ensemble neural networks

Abstract

Floods are among the most devastating natural disasters, causing significant loss of life and damage to essential infrastructure. The most effective way to mitigate flood impact is through early prediction and efficient post-flood management. However, one of the main challenges in flood monitoring is the lack of accurate and timely data, often due to adverse weather conditions. Synthetic Aperture Radar (SAR) imagery has provided a partial solution, as it can capture information regardless of weather or illumination. However, analyzing SAR images solely based on backscattering coefficients presents difficulties, especially in complex terrains affected by distortions. To address this, integrating Digital Elevation Models (DEMs) can enhance flood classification by accurately capturing terrain slopes. In this study, we propose a novel lightweight terrain-aware ensemble NN model that integrates Sentinel-1 VV and VH polarization bands with the Local Positive Neighborhood Averaged (LPNA-corrected) Copernicus DEM and terrain slope to enhance flood classification performance. The results demonstrate a relative improvement in classification accuracy, highlighting the effectiveness of multi-source data fusion with hierarchical NN for flood mapping.