Flood Detection from Satellite Images Using Self-Attention Empowered EfficientNetV2 Mechanism

Floods occur when water overflows onto normally dry land and are a destructive natural disaster. In recent times, deep learning models have demonstrated their remarkable capabilities in identifying objects and classifying images. As a result, the utilization of deep learning in remote sensing, especially for the crucial task of flood detection, has become an increasingly prominent and promising area of research and application. This study proposes a lightweight yet powerful feature extraction framework by leveraging a self-attention(SA) based EfficientNetV2. The EfficientNetV2 architecture serves as the backbone of the model and has exhibited effectiveness in image classification tasks. By incorporating self-attention mechanisms, the model can capture spatial relationships within remote sensing images, enhancing its accuracy and efficiency in flood detection. The model was trained and assessed using the SEN12-FLOOD dataset, and the outcomes demonstrate that the suggested framework exhibits enhanced flood detection performance, achieving an accuracy of 95.9%.