Benchmark for Building Segmentation on Up-Scaled Sentinel-2 Imagery

Currently, we can solve a wide range of tasks using computer vision algorithms, which reduce manual labor and enable rapid analysis the environment. The remote sensing domain provides vast amounts satellite data, but it also poses challenges associated with processing this data. Baseline solutions intermediate results are available for various tasks, such as forest species classification, infrastructure recognition, emergency situation Despite these advances, two major issues high-performing artificial intelligence algorithms remain in current decade. first issue relates to availability To train robust algorithm, reasonable amount well-annotated training data is required. second another concern. Even though there number providers, high-resolution up-to-date imagery extremely expensive. This paper aims address by proposing an effective pipeline building segmentation that utilizes freely Sentinel-2 10 m spatial resolution. approach use combines super-resolution (SR) component semantic component. As result, simultaneously consider analyze SR improve quality through medium-resolution Additionally, collected made unique dataset Russian Federation covering area 1091.2 square kilometers. adjusted resolution 2.5 accompanied masks. footprints were created OpenStreetMap was manually checked verified. Several experiments conducted task, advanced image methods diffusion-based SR3 model, RCAN, SRGAN, MCGR. MCGR network produced best PSNR 27.54 SSIM 0.79. obtained images then used tackle task different neural models, including DeepLabV3 encoders, SWIN, Twins transformers. SWIN transformer achieved results, F1-score 79.60.