Near Real-Time Flood Mapping with Weakly Supervised Machine Learning

Advances in deep learning and computer vision are making significant contributions to flood mapping, particularly when integrated with remotely sensed data. Although existing supervised methods, especially convolutional neural networks, have proved be effective, they require intensive manual labeling of flooded pixels train a multi-layer network that learns abstract semantic features the input This research introduces novel weakly approach for pixel-wise mapping by leveraging multi-temporal remote sensing imagery image processing techniques (e.g., Normalized Difference Water Index edge detection) create labeled Using these data, bi-temporal U-Net model is then proposed trained detection without need time-consuming labor-intensive human annotations. floods from Hurricanes Florence Harvey as case studies, we evaluated performance baseline models, such decision tree, random forest, gradient boost, adaptive boosting classifiers. To assess effectiveness our approach, conducted comprehensive assessment (1) covered multiple test sites varying degrees urbanization, (2) utilized both (i.e., pre- post-flood) uni-temporal only input. The experimental results showed framework data generation could produce near real-time urban maps consistently high precision, recall, f1 score, IoU overall accuracy compared machine algorithms.