ABSTRACT:

Sub-meter DBM were collected using amphibious UAVs to establish a new database. A topographic and spectral constrained Transformer GAN (TSCTGAN) was developed for 4× SR reconstruction, which incorporates a lightweight Transformer attention mechanism, multi-physical constraints and dynamic weighting strategy based on homoscedastic uncertainty to enhance reconstruction accuracy. Results demonstrate that the performance of TSCTGAN outperforms other SR techniques such as ESRGAN, DDPM, and TfaSR. Compared to the Bicubic method, TSCTGAN achieves a 3.39 dB improvement in PSNR and reduces RMSE from 0.489 m to 0.291 m. Moreover, the model exhibits robust generalization capabilities on cross-equipment and cross-scale extrapolation datasets.