Fusion of SAR and Optical Data with Flood Susceptibility Model for Mapping Floods and Assessing Impacts in Rural Area

Remote rural regions frequently experience severe flooding due to the high frequency and intensity of rainfall events. Prompt evaluation of these effects is essential for initiating relief efforts and reducing risks in susceptible areas. Nonetheless, the limited availability of data and the inherent constraints of conventional assessment methods present substantial obstacles. Typically, traditional change detection workflows are intricate and demand considerable human involvement. In contrast, deep learning presents an alternative that necessitates minimal human input and a reduced amount of labelled data. This study aims to evaluate and quantify the impacts of flooding by amalgamating Sentinel-1 synthetic aperture radar (SAR) data and Sentinel-2 optical imagery with a flood susceptibility model. The methodology encompasses a three-phase approach: delineating the extent of flooding, comparing classification accuracy through a novel deep learning-based change detection classifier, and evaluating flood vulnerability via Geographic Information Systems (GIS)-based modelling. Two attention mechanisms, namely the Pyramid spatial-temporal attention module (PAM) and Basic spatial-temporal attention module (BAM), were trained and evaluated using ResNet architectures. Findings indicate that PAM combined with ResNet 34 exceeded the performance of BAM with ResNet 34, achieving a remarkable overall accuracy of 98.80% and a Kappa Coefficient of 96.80%. In a separate trial utilizing a ResNet 18 backbone, the PAM attention mechanism yielded superior overall accuracy, attaining Kappa coefficients of 98.40% and 96.20%, respectively. Notably, the performance of the deep learning model exceeded that of traditional change detection methods, which recorded accuracies of 92.5% and 90.80%. To assess the vulnerability of the region and evaluate flood impacts, a flood susceptibility model was generated using a GIS-based Multi-Criteria Decision Model Analysis (MCDA). The weighted overlay technique in ArcGIS Pro 3.5 was employed to create flood susceptibility maps, categorised into five levels: very low, low, moderate, high, and very high. A significant portion of the area was identified within the high to very high susceptibility categories, with 453.37 km² (26.79%) classified as very high susceptibility and 488.93 km² (28.89%) as high susceptibility. The flood risk maps demonstrated a 97.44% agreement with the flood extent derived from Sentinel-1 imagery. The synergistic application of the developed flood risk map alongside flood extent mapping provided critical insights for spatial analysis and evaluation of flood impacts in Ingham and throughout the Shire of Hinchinbrook, Queensland, Australia.