Improving water and land classification in SWOT satellite pixel cloud data using deep learning : a case study in northeastern Germany

Abstract

Accurate water resource monitoring is vital for supporting human life, economic activities, and ecosystem stability. While the Surface Water and Ocean Topography (SWOT) satellite provides valuable observational data for water monitoring and management, its original water-land classifications contain misclassified points that compromise measurement accuracy. To overcome this problem, Deep Neural Network (DNN) models are developed to perform precise binary water-land classification using SWOT pixel cloud data. The development of a comprehensive deep learning framework begins with a baseline Single Point Features (SPF) model and advances to three optimized architectures: Nearest Heights (NH), Nearest 3 Features (N03F), and Convolutional Nearest Features (CNF) models, which strategically incorporate spatial features and multi-feature combinations. However, each DNN model possesses both advantages and disadvantages, precluding the selection of a single optimal model. Consequently, a majority voting ensemble strategy is implemented, with tuning of voting threshold P to optimize observational accuracy and data quantity. Following model training, DNN models are assessed across 14 test epochs using metrics: (1) Root Mean Squared Error (RMSE) for accuracy of river profiles, (2) Interquartile Range (IQR) for height stability of water points on lake, and (3) Intersection over Union (IoU) for lake coverage assessment. In case study of Peene river, the ensemble model demonstrate strong improvement, reducing median RMSE of river profiles by 43.62% (from 2.43m to 1.37m) compared to river profiles derived from official riverSP products. In case study of Dolgener See, the ensemble model simultaneously achieves better IoU value compared to "all water-related categories" of original SWOT classifications, while maintaining an IQR value similar to "open water" category in original SWOT classifications. In conclusion, this study demonstrates the DNN's capability to enhance SWOT-based water monitoring through precise binary classification. The ensemble of DNN effectively balances the accuracy and coverage of water observations.