RU-Net++: An automatic extraction method for Impervious Surface Area based on neural networks

Abstract. Impervious Surface Area (ISA) is vital for urban planning, environmental monitoring, and water management. Traditional remote sensing methods struggle with complex urban landscapes, leading to accuracy limitations. To address this, we propose RU-Net++, a deep learning-based ISA extraction model integrating ResNet50 as the encoder with spatial, channel, and dual attention mechanisms. The decoder employs an Atrous Spatial Pyramid Pooling (ASPP) module and multiple refinement modules to enhance feature representation and edge restoration. Trained on GLC_FCS30D and GISA datasets, RU-Net++ outperforms traditional methods in IoU, F1 Score, and Overall Accuracy, offering a reliable tool for sustainable urban development and land-use management.