Answers are grounded in this paper and cite their sources.
Optical image-guided digital surface model (DSM) super-resolution reconstructs fine-grained elevation data by leveraging coregistered high-resolution optical images. Existing methods, however, suffer from three critical limitations: first, the lack of semantic-level cross-modal alignment causes fusion misregistration in regions that look similar yet differ in elevation; second, indiscriminate transfer of textures from the optical image introduces spurious elevation fluctuations over flat surfaces, a problem known as texture leakage; finally, existing attention mechanisms fail to exploit the spatially nonuniform nature of structural information in DSMs. To address these challenges, we propose SemGeoNet, a semantic–geometry fusion network that establishes synergistic cross-modal interaction at three complementary levels. The Prototypical Semantic Context Module constructs a shared semantic space via learnable prototypes for category-level alignment. The Semantic-Guided Detail Rectification Module generates spatial quality maps that selectively transfer geometry-relevant details while suppressing texture leakage. The gradient-prioritized sparse attention mechanism uses elevation gradients as priors to guide global attention toward topographically salient regions for efficient long-range modeling. The three modules are integrated within a dual-stream encoder–single-stream decoder architecture. Experiments on the Vaihingen and Cologne datasets at scale factors of × 4 , × 8 , and × 16 show that SemGeoNet outperforms state-of-the-art methods across root-mean-square error (RMSE), mean absolute error, peak signal-to-noise ratio (PSNR), and structural similarity index. At × 16 , SemGeoNet achieves RMSE reductions of 38.5% and 11.6%, and PSNR gains of 4.22 and 2.42 dB, over the second-best method on the two datasets, respectively.
No takes yet. Share an insight, caveat, or question.
Song et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: