Improving Satellite Flood Image Classification Using Attention-Based CNN and Transformer Models

Floods are among the most frequent and devastating natural disasters, significantly impacting infrastructure, ecosystems, and human communities. Accurate satellite-based flood image classification is crucial for assessing flood-affected regions and supporting emergency response efforts. This study uses Convolutional Neural Networks (CNNs) and transformer-based architectures to enhance flood classification, integrating the Convolutional Block Attention Module (CBAM) to improve feature extraction. Using the xView2 xBD dataset, we classify houses as completely or partially surrounded by flood-water. Experimental evaluations demonstrate that ResNet101v2 achieved an accuracy of 86.87%, while a hybrid CNN model (MobileNetV2- DenseNet201) attained 85.83%, further improving to 89.54CBAM. The Vision Transformer (ViT) with CBAM achieved the highest accuracy of 90.75%, showcasing the effectiveness of attention-based hybrid models for flood image classification. These results highlight the potential of integrating CBAM with deep learning architectures to enhance classification accuracy and improve flood impact assessment.