Author(s): Haoran Li; Chenxi Ma; Zecong Tang; Boyuan An; Chao Qin; Yuan Xue; Ziyi Wang; Yicheng Ma; Xudong Fu
Linked Author(s):
Keywords: River cross section morphology hydrological static covariates deep learning interpretable artificial intelligence comprehensive evaluation system
Abstract: The morphology of river cross sections forms the foundation for studying river hydrological processes and material fluxes. The acquisition of cross section shapes relies primarily on field measurements, which restricts the ability to obtain data in remote regions. Multi-source remote sensing technology, employing integrated space-based systems, enables the large-scale extraction of river cross section shapes above the lowest water level. However, limited studies have addressed measurement and prediction methods for cross section morphology below the lowest water level. Using 568 measured sections and the Temporal Fusion Transformer (TFT) deep learning model, a prediction framework for underwater cross section morphology was developed, utilizing terrain and hydrological static covariates above the lowest water level. A comprehensive evaluation framework was designed to assess the model's performance. Visualization of the neural network module identified the key factors influencing cross section formation. The primary findings are as follows: (1) The TFT deep learning model demonstrates significant potential in predicting the underwater cross section morphology of single thread rivers, achieving a relative mean square error (RMSE) of 0.254 on the test set; (2) Static covariates, including climate type, annual mean temperature, potential evaporation, and annual mean runoff, are the main factors influencing river cross section morphology. The research findings provide insights and technical support for the automatic, systematic, and detailed extraction of river cross section morphology and related information in data scarce regions or large basins.
Year: 2025