MAPSRNet: a task-oriented super-resolution network that enhances building extraction beyond raw high-resolution imagery

Super-resolution (SR) is essential for remote sensing applications where high-resolution (HR) imagery is unavailable or difficult to acquire. Although many existing SR networks achieve high reconstruction performance in terms of the peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM), these improvements do not necessarily translate into better performance in downstream remote sensing tasks. In particular, when SR images are used for building footprint detection, the performance gains are often limited. This is because PSNR and SSIM mainly evaluate pixel-level fidelity and visual similarity, which favor accurate color and texture reconstruction but do not explicitly ensure the recovery of structural details such as object boundaries and geometry that are critical for segmentation tasks. To address this limitation, this paper proposed a multi-attention pyramid super-resolution network (MAPSRNet), which integrates spatial attention, cross-channel attention, and self-attention mechanisms to enhance the recovery of both visual details and structural information. By jointly modeling these complementary attention mechanisms, MAPSRNet aims to preserve building geometry and spatial structures that are important for downstream segmentation. Experiments are conducted on three public building datasets, Massachusetts (MAS), Wuhan University (WHU), and Waterloo (WAT). Segmentation results obtained using UNet, SAMv3, and ConvNeXtV2 demonstrate that MAPSRNet consistently produces SR images that are more beneficial for building detection than low-resolution inputs and outputs generated by representative SR methods. On the MAS dataset, MAPSRNet-based inputs improve building detection over the LR baseline by +3.61% in IoU with UNet, by +16.44% in IoU with SAMv3, and by +5.89% in IoU with ConvNeXtV2. On the WHU dataset, MAPSRNet-based inputs improve building detection over the StableSR baseline by +5.04% in IoU with UNet, over the SRResNet baseline by +11.4% in IoU with SAMv3, and over the SupResDiffGAN baseline by +4.82% in IoU with ConvNeXtV2. In some cases, MAPSRNet-based inputs can even match or surpass HR-trained baselines; for example, on the WAT dataset with ConvNeXtV2, the IoU reaches 72.53% compared with 71.84% for HR, while the F1-score improves to 84.08% compared with 83.61% for HR. These results demonstrate that MAPSRNet generates task-aware SR reconstructions that better preserve structural information relevant to building segmentation, providing an effective SR solution for urban mapping and large-scale remote sensing applications.