Publication

Abstract : Accurate road extraction from satellite imagery is a critical task in remote sensing and geospatial analysis, enabling applications such as autonomous navigation, urban planning, and disaster response. Despite recent advancements in deep learning, effectively capturing and emphasizing relevant features for road segmentation remains a significant challenge due to complex backgrounds and varying road conditions. In this work, we introduce a Feature-Selective U-Net architecture tailored for road extraction tasks. This architecture integrates a standard U-Net backbone with a novel Feature Selection Module based on the Convolutional Block Attention Module (CBAM), allowing the model to dynamically recalibrate feature importance by enhancing informative patterns and suppressing irrelevant signals. We evaluate our approach on two benchmark datasets-the Massachusetts Roads Dataset (MIT) and the DeepGlobe Road Extraction Dataset (DG)-achieving notable gains in Intersection over Union (IoU) of 3.5% on MIT and 0.41% on DG compared to UperNet. These results demonstrate the robustness and effectiveness of our method in diverse satellite imaging conditions. The code will be made publicly available at: https://github.com/amtom2004/cbam_fsu_roadseg