Cost-Effective River Bathymetry Reconstruction: Advanced Measure Strategies and Deep Learning

The river morphology data (bathymetry) is crucial for riverine simulations, including flood modeling, ecological assessments, and sediment transport, among others. Bathymetry is mainly measured by single- and multi-beam boat-mounted echosounders, and previous studies have investigated methods to utilize multi-beam and cross-sectional single-beam data. However, irregular or zigzag single-beam data processing methods and configurations remain complicated and unclear. Besides, bathymetry is precious due to practical limitations of equipment, plans, and budgets. Therefore, this dissertation explores approaches to bathymetric reconstruction using single-beam surveys and deep learning (DL) models through three research objectives. Specifically, the three research objectives are (i) to investigate cost-effective single-beam survey configurations, (ii) to develop a DL framework for river channel cross-section prediction, and (iii) to enhance the DL techniques for more general (bathymetric raster) prediction. The three research objectives of this dissertation are to comprehend people’s understanding of river bathymetry.

The first objective is to find out cost-effective bathymetric survey configurations considering accuracy, cost, and interaction with interpolations. The results indicate that cross-sectional surveys, while requiring fewer survey paths, effectively capture a river’s lateral variability. The study also establishes cost-effective spacing and wavelength recommendations for various river types. It demonstrates that single-beam surveys can achieve errors below 10% of the maximum depth, offering an economical alternative to multibeam surveys.

The second and third objectives focus on addressing data scarcity challenges by leveraging DL techniques to generate synthetic bathymetry. Advances in DL methods have enabled them to outperform conceptual and physical models in terms of accuracy, computational efficiency, and input accessibility. In the second objective, a Conditional Generative Adversarial Network (CGAN) is developed to generate synthetic river cross-sections. By integrating physics-informed input parameters and normalization, the CGAN model achieves a normalized root mean square error (RMSE*) of approximately 0.25 in vertex elevations, lower than the error resulting from conceptual models (0.45). Moreover, using normalized reach properties, the model can work across rivers on various scales. Despite some limitations and assumptions, this study underscores the potential of DL models in hydraulic and hydrologic applications.

The third objective extends deep learning to two-dimensional bathymetry prediction using a GAN-based Recurrent Feature Reasoning Network (RFR-Net). This high-resolution image inpainting technique is applied to fill hydro-flatten areas within digital elevation models (DEMs). Unlike the CGAN model in the second objective, which relies on manually extracted reach properties based on domain knowledge, the RFR-Net autonomously weighs and prioritizes critical features through the learning process. The model mostly maintains RMSE* values below 1.0 and under 0.5 across test sites in some cases. Although this approach offers a more generalized and automated prediction framework, it still suffers from data quality and quantity challenges. Enhancing model performance and applicability will require more robust cyberinfrastructure and expanded datasets.

This research advances both survey planning and deep learning applications for river bathymetry reconstruction. The proposed methodologies improve data efficiency, reduce costs, and enhance predictive accuracy across diverse river systems, offering valuable insights for hydrological and environmental modeling.