Dual-Stream Input Gabor Convolution Network for Building Change Detection in Remote Sensing Images

Accurate positioning and reduction of false change detection are the main requirements of building change detection in remote sensing images. In order to reduce the appearance of pseudo-changes, this paper proposes a building change detection model based on improved UNet++ using Siamese architecture. The model uses Gabor convolution to mine the spatio-temporal relationship between bi-temporal images from multi-direction and multi-scale to achieve pixel level semantic correction. In the output part, the feature fusion module is used to fuse the multi-scale output, avoiding the semantic confusion caused by direct fusion. In order to validate the proposed model, comparison experiments were conducted with FC-Conc, SNUNet, STANet, ESCNet and other models on LEVIR-CD and WHU-CD datasets. The experimental results show that the F1 score and IOU of the proposed model have reached the best performance. In addition, ablation experiments were performed. Compared to the baseline model UNet++, the F1 scores have improved by 5.9% and 1.02% respectively, while the IOU have increased by 9.24% and 1.47% respectively.