A fast approach for flood mapping over a large region using Sentinel-2 imagery

Abstract

Floods impact millions globally, necessitating rapid mapping techniques for effective disaster management response. This study introduces a novel patch-wise classification framework and the ChangeSEN2-FL dataset, comprising 701 flood and 1216 non-flood multispectral bitemporal patch pairs from diverse global regions using Sentinel-2 imagery. Each 5km x 5km patch includes 12 pre-event and 12 post-event bands. Using the ChangeSEN2-FL dataset, we developed a classification framework with improved generalization, utilizing multiple normalized difference indices (NDIs) for reliable flood detection. To address flood variability, we introduce a multiple thresholding approach (MTA) for NDIs within a dual-input ResNet50-based architecture. This setup utilizes false-colour (FC) and true-colour (TC) RGB composites to improve flood detection generalizability. The proposed framework is tested on a held-out set and three case studies from Madagascar, Pakistan, and Australia. The case studies test the global applicability of the model framework, comparing performance across regular RGB and dual-input configurations of single and multiple thresholded false-colour RGB. The held-out test results show a 3% improvement over TC RGB-based model, achieving 97% accuracy. Case studies further confirm adaptability, with 7.5% and 55% improvement over TC RGB-based model in Madagascar and Pakistan, respectively, demonstrating strong generalization across diverse flood scenarios.