One of the most damaging natural disasters, floods pose a serious risk to human life, infrastructure, and economies. Effective disaster response depends on the timely and precise identification of areas affected by flooding. A deep learning-based semantic segmentation framework designed for flood scene analysis with satellite or drone imagery is presented in this paper which facilitates speedy flood response and post-event mapping for successful disaster management. The input images are preprocessed using bilateral filtering and Contrast Limited Adaptive Histogram Equalization (CLAHE) to improve visual clarity and class separability. The segmentation model has a lightweight custom decoder for accurate pixel-wise classification and an encoder based on Swin Transformer for capturing hierarchical features. Training and evaluation are conducted using the FloodNet dataset, which comprises ten flood-relevant semantic classes. To enhance generalization and further refine spatial boundaries, postprocessing techniques like Test-Time Augmentation (TTA) and Dense Conditional Random Fields (DenseCRF) are used. Compared to traditional CNN-methodologies, the proposed approach has been able to achieve a validation mean Intersection over Union (mIoU) as high as 84.1% at 224 × 224 input resolution with DenseCRF when dealing with visually complex flooding conditions.
Preetha et al. (Mon,) studied this question.