An applicable and automatic method for earth surface water mapping based on multispectral images

Earth’s surface water plays an important role in the global water cycle, environmental processes, and human society, and it is necessary to dynamically capture the distribution and extent of surface water on Earth. However, due to the high complexity of the surface environment of Earth, the current surface water mapping methods are limited in applicability and precision. In this study, to explore an automatic and applicable model for surface water mapping, particularly for the regions with highly heterogenous backgrounds, we adopted state-of-the-art deep learning techniques and structured a new model, namely, WatNet, for surface water mapping. Specifically, we combined a state-of-the-art image classification model and a semantic segmentation model into an improved deep learning model. For the fine-scale identification of small water bodies, the combined model was further improved with surface water mapping-tailored design. To learn the surface water features of worldwide regions, a surface water knowledge base that consists of worldwide satellite images was built in this study. The newly structured WatNet model was tested on three highly heterogeneous regions, and as demonstrated by the results, 1) the trained WatNet model achieved the highest accuracies, which were above 95%, for all the selected test regions; 2) the new structured WatNet model yields significant improvements through state-of-the-art model combinations and the surface water-tailored design; and 3) unlike conventional methods, which usually require parameterization in accordance with the specific surface environment, trained WatNet can be directly applied for highly accurate surface water mapping, and, thus, no human labor is required.