Author(s): Haoran Li; Chenxi Ma; Zecong Tang; Rong Li; Zixu Wang; Boyuan An; Chao Qin; Yuan Xue; Ziyi Wang; Xudong Fu

Linked Author(s): Haoran LI

Keywords: River cross section morphology; Deep learning; Domain generalization; Quantile prediction

Abstract: River cross section morphology provides critical boundary conditions for hydrological simu-lations and fluvial geomorphology studies. Traditional methods in river cross section mor-phology acquisition exhibit significant limitations in capturing heterogeneous underwater topographic features, particularly in mountainous rivers characterized by turbulent flow re-gimes and restricted accessibility. Although deep learning techniques such as Temporal Fu-sion Transformers (TFT) offer new opportunities for cross section reconstruction, existing methods still face challenges in integrating multi-source data and achieving cross-domain generalization. To address these issues, this study proposes a novel training strategy by in-tegrating contact-based field measurement data with high-resolution unmanned aerial vehi-cle (UAV) remote sensing imagery to achieve precise reconstruction of complex river cross section morphology. The strategy incorporates three key innovations: (1) a quantile-based probabilistic prediction framework to quantify uncertainty boundaries in cross section mor-phology; (2) a confidence-guided adaptive annotation enhancement algorithm leveraging multi-scale models (e. g., a Loess Plateau-specific model) to optimize sparse UAV remote sensing data quality; and (3) a domain generalization training strategy to mitigate cross-domain distribution shifts. Experimental results demonstrate that both progressive data ratio training and two-stage sequential training achieve robust performance on test sets (R2 > 0.45). This work provides a robust technical foundation for integrating hydrological data from localized high-precision measurements to full-coverage applications, advancing the development of digital twin river basins.

DOI: https://doi.org/10.64697/978-90-835589-7-4_41WC-P1753-cd

Year: 2025