Multiple spatial and wavelet feature interaction learning for change detection in remote sensing images

Change detection (CD) in remote sensing (RS) images is essential for earth observation, monitoring dynamic environmental and urban changes. Existing deep learning (DL)-based methods often struggle with pseudochanges and subtle small-scale changes. To address this, we propose a spatial and wavelet interactive learning (SWIL) framework, which jointly optimizes spatial and spectral feature learning. By leveraging multidomain feature complementarity, SWIL suppresses pseudochanges while enhancing fine-scale detection, improving noise robustness. Specifically, we integrate the wavelet transform into feature filtering, exploiting its multiscale geometric analysis for fine-grained frequency-domain representation. Our spatial and wavelet feature enhancement (SWE) module decomposes and aggregates multiscale features with local convolutions, effectively reducing pseudo-change noise. Additionally, the Channel Selective Multiscale Filtering (CSF) module refines spatial-frequency features, improving small-change detection in complex scenes. Experiments on three benchmarks confirm SWIL's superiority over state-of-the-art methods, achieving higher accuracy and noise resilience under complex conditions.