Super-Resolution of Remote Sensing Images via the Parallel Lattice Attention Network

Deep learning-based super-resolution models often struggle to balance capturing long-range dependencies with computational efficiency, as convolutional neural networks focus on local features while transformers incur high computational costs. To address this trade-off, we propose a parallel lattice attention network (PLAN) to improve the reconstruction of complex features. Unlike traditional lattice networks that rely on sequential cascading, PLAN introduces a parallel lattice attention block that splits the input feature map into concurrent branches. These branches employ lattice attention units to simultaneously extract fine-grained details and global context, which are subsequently fused via attention-based integration. This parallel design maximizes computational efficiency while ensuring robust reuse of features. Experiments on the AID, UC Merced, and WHU-RS19 datasets at various scales show PLAN's superiority in focusing on critical information to enhance image quality. On the UC Merced dataset with a$2 \times$scaling factor, PLAN achieves a PSNR of 32.12 dB and an SSIM of 0.8807, outperforming state-of-the-art methods such as ESTNet and BMFENet by at least 0.07 dB in PSNR and 0.0210 in SSIM. On the WHU-RS19 dataset with a$4 \times$scaling factor, PLAN achieves a PSNR of 29.16 dB and an SSIM of 0.7597, surpassing previous methods by up to 0.64 dB in PSNR. PLAN also maintains computational efficiency with a runtime of 30.97 ms, faster performance than ESRT (106.34 ms), SwinIR (98.47 ms), FENet (63.52 ms), OmniSR (45.02 ms), BMFENet (51.43 ms), MSCT (57.15 ms), and ESTNet (67.95 ms).