ASPCD-UNet: An Improved Network for Change Detection

Change detection (CD) is used to identify the required significant changes between two-phase images. High-resolution remote sensing image CD plays an important role in land and resources planning, military strategy research, geological exploration research, natural disaster prediction and other fields. Given two identical registration photographs taken at various times, lighting or seasonal changes are frequently present, as well as other issues such significant changes in the image of the real thing. Despite these issues, the real object's small size and its boundary changes are still a concern. Therefore, we propose a Siam network based on UNet to solve some subtle problems in CD for edge information and small object detection. The original convolutional module in our UNet network is improved into an optimization module in the form of jumper connections, a pyramidal pool module is added at the bottom of the encoder to better extract global information, and these changes together create an improved ASPCD-UNet network model. The model training is based on cross-entropy loss function, with a variety of optimization strategies to achieve end-to-end ground object change information extraction.