Abstract: The deep learning-based change detection of remote sensing images has seen rapid advances in the past few years. However，it still faces challenges for change detection in complex scenes，such as incomplete recognition and high false detection rates. In response to these challenges，this paper proposed the FTUNet，a network based on SNUnet that integrates the fast Fourier transform （FFT） and efficient multi-head self-attention （EMHSA）. Specifically，the FFT module in the network enabled style unification of dual-temporal images，reducing false detection caused by “pseudo changes” due to external factors such as light variations. Additionally，the EMHSA was introduced in the feature extraction stage to fully extract the contextual information from the feature maps，thereby enhancing the segmentation integrity of target changes. Experiments on the LEVIR-CD and SYSU-CD public datasets showed that the FTUNet exhibited increases of 1.42 and 1.53 percentage points in F1 score，as well as increases of 2.31 and 2.07 percentage points in intersection over union （IoU），compared to the SNUNet.