INTERGEO Agenda - english

In a changing climate, the need to monitor and manage water resources presents a key challenge. The launch of the new Surface Water and Ocean Topography (SWOT) mission marked the start of a new era in satellite altimetry, enabling unprecedented coverage of Earth’s freshwater resources. With this enhanced spatial coverage and the very high spatial resolution of the pixel cloud (PIXC) product, both facilitated by the innovative, wide-swath measurement system on board the satellite, new possibilities in global water body monitoring are emerging. In this study, we present a deep-learning-based approach to enhance water detection in SWOT PIXC data. The proposed method employs a dynamic graph convolutional architecture to incorporate both spatial proximity and feature similarity to produce binary class label predictions under highly imbalanced conditions. The model is trained on a full year of SWOT PIXC data covering the Dallas-Fort Worth metropolitan area in the United States of America with contemporary class labels from the DSWx product suite, which are matched at the pixel level. Subsequently, the model is validated on an independent test set generated from the subsequent year and compared to the native classification attribute provided in the SWOT PIXC data product. Against DSWx-derived ground truth labels, the proposed method improves binary water-land classification from a scene-level mean F1 score of 0.597 to 0.913 on the temporally disconnected test set. The emphasis of the study lies in evaluating the effectiveness of the proposed method in land-water delineation at the PIXC level. A particular focus is on the mitigation of misclassifications related to high-powered radar returns caused by built surfaces as a predominant source of misclassification in urban environments. The results provide insights into the suitability of dynamic graph neural network-based approaches for land-water classification at the highest spatial resolution that the groundbreaking SWOT mission provides. Improvements offer the prospect of more accurate flood monitoring and management, in particular, in urban environments.