| Abstrakt: |
This paper examines the multi-scale super-resolution challenge of digital elevation models in remote sensing. A dual-domain multi-scale attention fusion network is proposed, which reconstructs digital elevation image details step-by-step using cascading sub-networks. This model incorporates components like the wavelet guidance and separation module, multi-scale attention fusion blocks, dilated convolutional inception module, and edge enhancement module to improve feature extraction and fusion capabilities. A new loss function is designed to enhance the model's robustness and stability. Experiments indicate that the proposed model outperforms 15 benchmark models in PSNR, RMSE, MAE, R M S E s l o p e , and R M S E a s p e c t metrics. In HMA data, The proposed model's PSNR increases by 0.89 dB (~1.81%), and RMSE decreases by 1.22 m (~8.6%) compared to a state-of-the-art model. Compared to EDEM, which has the best elevation index, R M S E s l o p e decreases by 0.79 ° (~16%). Additionally, the effectiveness and contribution of each DDMAFN component were verified through ablation experiments. Finally, on the SRTM dataset, The proposed model demonstrates superior performance even with interpolated degradation. [ABSTRACT FROM AUTHOR] |
