An End-to-end Framework For Low-Resolution Remote Sensing Semantic Segmentation
This addresses the challenge of limited access to high-resolution satellite imagery for remote sensing applications, though it is incremental as it builds on existing super-resolution and segmentation methods.
The paper tackles the problem of performing semantic segmentation on low-resolution remote sensing images by proposing an end-to-end framework that combines super-resolution and segmentation modules, achieving performance close to high-resolution data and surpassing low-resolution baselines.
High-resolution images for remote sensing applications are often not affordable or accessible, especially when in need of a wide temporal span of recordings. Given the easy access to low-resolution (LR) images from satellites, many remote sensing works rely on this type of data. The problem is that LR images are not appropriate for semantic segmentation, due to the need for high-quality data for accurate pixel prediction for this task. In this paper, we propose an end-to-end framework that unites a super-resolution and a semantic segmentation module in order to produce accurate thematic maps from LR inputs. It allows the semantic segmentation network to conduct the reconstruction process, modifying the input image with helpful textures. We evaluate the framework with three remote sensing datasets. The results show that the framework is capable of achieving a semantic segmentation performance close to native high-resolution data, while also surpassing the performance of a network trained with LR inputs.