Key points are not available for this paper at this time.
Deep learning has been widely used in mapping small water bodies (SWBs) with very high-resolution PlanetScope imagery. Traditional convolutional neural network (CNN) and Transformer hybrid networks, which rely on single-scale patches for feature embedding at each stage, often introduce artifacts at feature boundaries and compromise their structural integrity. Moreover, current models do not adequately account for the shape complexity of small water bodies, which is a critical factor in mapping performance. To address these limitations, this paper proposed a hybrid multi-perspective transformer-based U-Net (MTU-Net). MTU-Net synergistically learns global context by extending single scale to multiscale embedding with multiple path Transformer. In the feature fusion stage, MTU-Net fuses shallow and deep features through spatial and channel attention and considers the water index to preserve rich spatial details and semantic information simultaneously. Furthermore, MTU-Net employs a shape-area adaptive-weighted binary cross entropy loss function, which dynamically adjusts sample weights during training by simultaneously considering the shape complexity and size of water bodies. Quantitative experiments demonstrate that MTU-Net achieves the highest accuracies F1 of 94.11% and an IoU of 88.87% compared with the state-of-the-art networks. The MTU-Net code is available: https://github.com/PUZHOU123/MTU-Net. The surface water mapping result in the Hubei Province, China is available under request.
Zhou et al. (Thu,) studied this question.