FDFormer: frequency difference-guided transformer for change detection in heterogeneous remote sensing images

Abstract Heterogeneous change detection (HCD) is a current hot topic in the field of remote sensing, as it can capture changed areas at the same geographical location using bi-temporal remote sensing images from different sensors. The primary challenge in HCD stems from the inherent domain shift between heterogeneous images, which fundamentally prevents the use of pixel-wise relationships and comparison analysis for change identification. Current advanced HCD methods have proposed deep learning-based modality alignment or modality-invariant feature extraction schemes to mitigate domain shifts between heterogeneous images and achieve effective HCD. However, current methods largely overlook the potential of frequency-domain representations for establishing reliable correspondences and modeling cross-modal relationships, creating a fundamental performance bottleneck in HCD. To overcome these limitations, we propose a frequency difference-guided Transformer (FDFormer) for HCD, which is composed of a frequency consistency sample generation (FCSG) strategy and a frequency difference representation Transformer (FDRT). With only 1% of pixels initially labeled, our FCSG strategy leverages local frequency consistency to select reliable pseudo-labels and mitigate modality discrepancies in the frequency domain. In addition, the proposed FDRT introduces a wavelet frequency difference-guided Transformer block to enhance the representation of difference features between heterogeneous remote sensing images by modeling the relationship between wavelet features of different frequencies between heterogeneous images. Ablation studies demonstrate the effectiveness of the proposed FCSG and FDRT. Furthermore, extensive comparative experiments on three public HCD datasets validate that the proposed FDFormer presents certain superiority and competitiveness compared to several state-of-the-art HCD methods.