Building Change Detection Based on Fully Convolutional Network in High-Resolution Remote Sensing Images

To deal with the issues of rough boundaries of varying objects and loss of tiny objects in building change detection, we design a fully convolutional-based building change detection method (BFFGNet). Initially, to capture the fine difference features at various scales, the Feature Difference Enhancement (FDE) module is proposed for enhancing interaction of information between the bi-temporal features. Then, for extracting accurate boundary information and making the edges of altered areas clearer, we propose the Boundary Feature Compensation (BFC) module to make up for the boundary information that is lost due to network deepening, and the boundary enhanced multi-level characteristics are merged by the Multi-Scale Feature Aggregation (MSFA) module to generate change guide map that contains more semantic and detail information. Finally, the extracted change guide map is utilized as prior information for directing the distinct level feature integration using the Change Guide Module (CGM), which enhances model's capacity to identify complete buildings and small targets. To demonstrate the model's effectiveness, it was tested on two large remote sensing building change detection datasets, theWHU-CD and LEVIR-CD datasets. Study shows that compared to sub-optimal network, the IoUscores of BFFGNet on these two datasets are raised by 0.85% and 2.86%, respectively, and BFFGNet significantly outperforms the comparative state-of-the-art (SOTA) methods.