ABSTRACT:

Digital bathymetric models (DBMs) are fundamental terrain inputs for hydrodynamic modelling and flood-emergency analysis, but rapid acquisition of high-resolution bathymetry remains challenging in complex and hazardous waters. This study develops a super-resolution (SR) reconstruction framework integrating amphibious unmanned aerial vehicles (UAVs), sonar measurements, and deep learning. High-density field sonar observations were spatially regularized to construct the reference high-resolution (HR) DBM, whereas realistic low-resolution (LR) inputs were generated by spatially thinning the original sonar soundings before rasterization, forming measurement-derived LR–HR pairs for 4× SR reconstruction. A topographic and spectral constrained Transformer GAN (TSCTGAN) was then developed by integrating lightweight Transformer attention, geometric and spectral constraints, and adaptive loss weighting based on homoscedastic uncertainty. On the test dataset, TSCTGAN achieved a PSNR of 49.56 dB and an RMSE of 0.177 m. Compared with Bicubic and Kriging reconstruction from the same sparse observations, PSNR increased by 2.33–4.56 dB and RMSE decreased by 51.9%–54.0%. Three extrapolation datasets representing cross-sensor, cross-resolution, and geographically independent cross-site conditions yielded TSCTGAN RMSEs of approximately 0.140, 0.063, and 0.113 m. For full-scene reconstruction, TSCTGAN achieved a depth RMSE of 0.136 m, a pairwise overlap RMSE of 0.021 m, and a seam depth-increment RMSE of 0.046 m. These results demonstrate that the proposed framework can retain favorable bathymetric accuracy and spatial continuity during patch-based full-scene reconstruction from sparse field observations. The framework addresses the bathymetric reconstruction stage preceding hydrodynamic modelling and provides a potential approach for rapidly generating spatially detailed DBM terrain inputs.

ABSTRACT:

Digital bathymetric models (DBMs) are fundamental terrain inputs for hydrodynamic modelling and flood-emergency analysis, but rapid acquisition of high-resolution bathymetry remains challenging in complex and hazardous waters. This study develops a super-resolution (SR) reconstruction framework integrating amphibious unmanned aerial vehicles (UAVs), sonar measurements, and deep learning. High-density field sonar observations were spatially regularized to construct the reference high-resolution (HR) DBM, whereas realistic low-resolution (LR) inputs were generated by spatially thinning the original sonar soundings before rasterization, forming measurement-derived LR–HR pairs for 4× SR reconstruction. A topographic and spectral constrained Transformer GAN (TSCTGAN) was then developed by integrating lightweight Transformer attention, geometric and spectral constraints, and adaptive loss weighting based on homoscedastic uncertainty. On the test dataset, TSCTGAN achieved a PSNR of 46.39 dB and an RMSE of 0.177 m. Compared with Bicubic and Kriging reconstruction from the same sparse observations, PSNR increased by 2.33–4.56 dB and RMSE decreased by 51.9%–54.0%. Three extrapolation datasets representing cross-sensor, cross-resolution, and geographically independent cross-site conditions yielded TSCTGAN RMSEs of approximately 0.140, 0.063, and 0.113 m. For full-scene reconstruction, TSCTGAN achieved a depth RMSE of 0.136 m, a pairwise overlap RMSE of 0.021 m, and a seam depth-increment RMSE of 0.046 m. These results demonstrate that the proposed framework can retain favorable bathymetric accuracy and spatial continuity during patch-based full-scene reconstruction from sparse field observations. The framework addresses the bathymetric reconstruction stage preceding hydrodynamic modelling and provides a potential approach for rapidly generating spatially detailed DBM terrain inputs.