ABSTRACT Accurate and timely coastal flood zone mapping is essential for disaster management, urban planning, and climate adaptation in at‐risk areas. This project uses deep learning to semantically differentiate flood‐affected Chinese coastal regions using high‐resolution annotated images. Pixel‐by‐pixel water classification in RGB aerial images is done using an Efficient U‐Net architecture, which combines a pre‐trained EfficientNet encoder with a U‐Net decoder. The proposed model was trained and tested using the public Flood Area Segmentation dataset, which is assumed to represent Chinese coastal floods. Due to the short sample size, many data augmentation strategies were applied to increase model generalization. Experiments demonstrate the Efficient U‐Net provides good segmentation. Dice coefficient 0.88, F1 score 0.90, and Intersection over Union 0.82 were the final validation metrics. Qualitative research suggests that predicted flood masks match ground reality annotations. Results suggest deep neural networks might automatically and accurately track coastal floods. They also established a standard for multisensor research. Further research will examine how to use remote sensing to improve coastal flood detection and monitoring.
Lin et al. (Mon,) studied this question.