A Conditionally Parameterized Feature Fusion U-Net for Building Change Detection

The semantic richness of remote sensing images often presents challenges in building detection, such as edge blurring, loss of detail, and low resolution. To address these issues and improve boundary precision, this paper proposes CCCUnet, a hybrid architecture developed for enhanced building extraction. CCCUnet integrates CondConv, Coord Attention, and a CGAFusion module to overcome the limitations of traditional U-Net-based methods. Additionally, the NLLLoss function is utilized in classification tasks to optimize model parameters during training. CondConv replaces standard convolution operations in the U-Net encoder, boosting model capacity and performance in building change detection while ensuring efficient inference. Coord Attention enhances the detection of complex contours in small buildings by utilizing its attention mechanism. Furthermore, the CGAFusion module combines channel and spatial attention in the skip connection structure, capturing both spatial and channel-wise correlations. Experimental results demonstrate that CCCUnet achieves high accuracy in building change detection, with improved edge refinement and the better detection of small building contours. Thus, CCCUnet serves as a valuable tool for precise building extraction from remote sensing images, with broad applications in urban planning, land use, and disaster monitoring.