TFFENet: Temporal Difference Fusion and Frequency-Domain Feature Ehancement Network for Remote Sensing Image Change Detection

Remote sensing change detection(RSCD) is an important technique for monitoring geographical change. However, under the influence of illumination factors such as imaging angle difference and illumination conditions, it is still a difficult task to accurately identify the change region. In addition, the existing methods only use limited information from bitemporal images, which leads to the loss of details such as small targets and change boundaries, thus leading to false detection. To address these limitations, we propose a novel network termed temporal difference fusion and frequency-domain feature enhancement network (TFFENet). First, we extract multi-level hierarchical features from the Siamese hierarchical backbone. Subsequently, a temporal difference module (TDM) is designed to generate differential features to enhance contour information. Moreover, a global feature enhancement module (GFEM) is introduced to strengthen the modeling of global contextual information. To effectively extract and refine texture features while mitigating the impact of illumination variations, we employ a frequency feature refinement module (FFRM) that fuses spatial and frequency information for adaptive feature filtering. Finally, to fully integrate heterogeneous feature representations, a depth feature extraction module (DFEM) is proposed to aggregate inter-layer contextual information and capture complementary feature semantics. To verify the effectiveness of the proposed TFFENet, extensive experiments were carried out on three public RSCD benchmark datasets, namely LEVIR-CD, WHU-CD, and SYSU-CD. The results show that TFFENet has achieved the best F1 values on all datasets, which are 91.91%, 94.99%, and 82.63%, respectively. The results consistently show superior performance compared to existing state-of-the-art (SOTA) methods. Our code will be publicly available athttps://github.com/amazinga1234/TFFENet.