Post-disaster flooded region segmentation using DeepLabv3+ and unmanned aerial system imagery

Natural disasters, particularly floods, have become increasingly frequent and intense in recent times, posing significant threats to human lives and infrastructure, especially in developing countries. Efficient flood detection and damage assessment are critical for effective disaster response and recovery. This study applies the DeepLabv3+ model with UAS imagery to achieve precise flood area delineation. The DeepLabv3+ model employs an encoder-decoder architecture, integrating Atrous Spatial Pyramid Pooling (ASPP) and atrous convolutions to capture multi-scale contextual features while preserving spatial details. To evaluate its performance, the study experiments with various backbone architectures, including ResNet-18, ResNet-50, MobileNetV2, and Xception, under different configurations of downsampling rates (8 and 16) and atrous rates (8, 12, and 16). ResNet-50 proves to be the most effective backbone, achieving the optimal balance between segmentation accuracy and computational efficiency. The ASPP module enhances global and local feature extraction, while the decoder combines low-level spatial and high-level semantic features for precise pixel-wise segmentation. Experimental results reveal that the DeepLabv3+ model significantly enhances the detection of flooded regions and the delineation of flood extents, providing a reliable tool for real-time disaster management and contributing to improved flood management practices. This research offers valuable insights into leveraging deep learning models for enhanced disaster response in regions where rapid and accurate flood detection is crucial.