TPI-Net: Time-Phase Interaction Network Based on Learnable Wavelet Attention for Remote Sensing Change Detection

Remote sensing change detection in urban buildings faces challenges in recognizing building details, extracting texture structures, and suppressing interference from nonchange areas. To address these issues, we propose the Time-Phase Interaction Network based on Learnable Wavelet Attention (TPI-Net). The model first employs ResNet50 for feature extraction, followed by the Deformable Local-Global Attention Interaction Module (DLGAM), which enhances feature consistency and suppresses nonchange area interference. For feature fusion, we introduce an Adaptive Feature Interaction Fusion Unit (AFIF) that uses the Sobel operator to enhance edge features and dynamically adjusts fusion strategies across different scales and channels. To capture texture detail changes, the Wavelet Multiple Attention Module (WMAM) incorporates learnable wavelet transform filters, allowing adaptive learning of different frequency components. This enables dynamic attention to texture and edge features. Additionally, a Dual-Focus Attention Mechanism is applied to enhance spatial and channel feature extraction. Finally, we validate DLGAM on three datasets. LEVIR-CD, WHU-CD, and GZ-CD. and show that our model achieves state-of-the-art performance across multiple benchmarks.