DSM Accuracy Enhancement via Satellite Stereo Imagery Super-Resolution: Multiscale Attention and Coordinate Mapping Optimization

Contemporary stereo optical satellite imagery imposes inherent limitations on the resolution and elevation accuracy achievable in digital surface model (DSM) extraction. To mitigate these constraints, we introduce a novel stereo image pair super-resolution (SPSR) methodology. Existing super-resolution (SR) research primarily focuses on single images and fails to address the geometric model degradation in stereo pairs postprocessing. Furthermore, a rigorous comparison is absent between direct SR application on DSM versus DSM extraction from SR-enhanced stereo imagery. Our solution employs a dual aggregation transformer with multiscale attention (DATMAT) network, which effectively resolves challenges in multiscale information aggregation and computational redundancy for remote sensing imagery. Crucially, by establishing a geometric position mapping model between super-resolved and source imagery, we rectify the loss of geolocational fidelity post-SR. Extensive experiments on WV2, SV1, GF7, and ZY3 datasets demonstrate that DATMAT significantly outperforms representative and recent deep learning methods (e.g., HAT, DAT, and EDSR). Quantitatively, DATMAT improves PSNR by 1.60%–19.08% and SSIM by 0.73%–8.03% over peer models. In terms of DSM vertical accuracy, although the performance gap among deep learning methods is inherently narrow, DATMAT achieves the lowest elevation RMSE across all tested platforms, reaching an optimal 2.46 m on WV2 and 3.78 m on GF7. Compared to the strongest competing Transformer-based models, DATMAT further reduces the elevation RMSE and MAE, while consistently maintaining the highest concentration of low-magnitude errors. These results indicate that DATMAT provides superior geometric model recovery and cross-scene robustness, establishing a high-precision, low-cost framework for satellite-based DSM production, promoting the practical utility of submeter resolution geospatial mapping.