Tackling Dual Gaps in Remote Sensing Segmentation: Task-Oriented Super-Resolution for Domain Adaptation

Semantic segmentation of remote sensing images plays a crucial role in various applications, such as land cover mapping and urban planning. However, the performance of semantic segmentation models often degrades when applied to images from different domains or with varying spatial resolutions. In this paper, we propose a novel task-oriented super-resolution method for domain adaptation in remote sensing semantic segmentation. Our approach aims to adapt a segmentation model trained on high-resolution images from a source domain to perform accurately on low-resolution images from a target domain. We introduce a super-resolution network that learns to enhance the spatial resolution of the target domain images while simultaneously optimizing the segmentation performance of a pre-trained and fixed segmentation model. The super-resolution network is trained using a combination of losses, including a segmentation loss, a perceptual loss, and a contrastive loss, which together ensure that the adapted images are both visually similar to the source domain images and semantically consistent with the ground-truth segmentation masks. We evaluate our method on two challenging remote sensing datasets, ISPRS Potsdam and Vaihingen, and demonstrate significant improvements in segmentation accuracy compared to state-of-the-art domain adaptation techniques. Our approach achieves mean Intersection over Union (mIoU) scores of 0.523 and 0.567 on the Potsdam and Vaihingen datasets, respectively. The proposed task-oriented super-resolution method offers a promising solution for adapting semantic segmentation models to new domains and resolutions in remote sensing applications.