A Cross-Domain Object-Semantic Matching Framework for Imbalanced High Spatial Resolution Imagery Water-Body Extraction

Large-scale information pertaining to surface water bodies is crucial for activities such as flood monitoring. Deep learning algorithms have shown great potential in water-body extraction based on high spatial resolution (HSR) imagery. However, the current reliance on deep learning for HSR imagery water-body extraction necessitates a substantial quantity of manually labeled training samples. The variance in spatial resolution among images and the intricacies of scenes consistently pose challenges to the transferability of deep learning. Moreover, the number of pixels representing water bodies is typically lower compared to the number of background pixels. This imbalance in class prediction probabilities often limits the accuracy of water-body class predictions. In this article, we propose a cross-domain object-semantic matching (COM) framework for extracting water bodies from unlabeled high-resolution remote sensing imagery. The distinctions in spectra, shapes, and semantic distributions of water bodies across various domains create challenges for certain source domain samples to contribute positively to model training. Therefore, a sample semantic similarity matching (S3M) mechanism is devised. The proposed object-contextual perception network (OCPNet) models multiscale water-body features and object-contextual representations (OCRs), aiming to achieve a more accurate and comprehensive representation of surface water bodies. In addition, to prevent the training process from being dominated by easily transferred categories in the target domain, a weighted joint loss is designed to alleviate the imbalance of predicted probabilities and pixel numbers between water and nonwater bodies. Experiments on four public datasets of GID, CCF, LoveDA, and DeepGlobe demonstrate the effectiveness and generalization of our proposed framework.