Enhancing flood segmentation in aerial imagery using a dual-decoder efficient U-Net with CBAM and post-hoc explainable AI technique

Precise flood segmentation from high-resolution UAV imagery is essential for rapid disaster response, however it faces difficulties due to complex textures, blurred water boundaries, and weak model interpretability. The present research proposes a unique segmentation framework, DualEff-CBAMNet, which incorporates EfficientNet as the encoder, two decoders for multiscale feature recovery, and CBAM attention to improve spatial intensity. A hybrid of Dice and Binary Cross-Entropy loss is used to enhance boundary identification and deal with class imbalance. To enhance generalisation, we implement a sophisticated augmentation pipeline including random flipping, rotation, resizing, elastic deformation, and ClassMix. Moreover, we apply LIME (Local Interpretable Model-Agnostic Explanations) to demonstrate the influence of superpixels on class predictions, offering insights into both correctly classified and misclassified areas. The proposed model was evaluated using a UAV flood dataset and compared to popular architectures such as U-Net+ResNet, Attention U-Net, FCN, SegNet, DeepLabV3+, TransUNet, Gated-SCNN, FPN, and Swin Transformer. DualEff-CBAMNet attains superior performance with 94.67% recall, 94.72% precision, 94.56% F1-score, 93.36% Dice, 87.45% mIoU, and 94.89% accuracy, surpassing all baseline models. LIME-based study demonstrates the model’s efficiency in precisely identifying flood areas while detecting limitations due to water-shadow ambiguity. The system provides a comprehensive and elucidative solution for UAV-based flood segmentation, enhancing both model efficiency and interpretability in actual disaster monitoring conditions.