Heterogeneous change detection via frequency domain interaction and statistical style embedding

Combining the all-weather sensing of Synthetic Aperture Radar (SAR) with optical imagery enables reliable Earth observation, yet their fundamentally different imaging mechanisms cause large representational discrepancies that hinder feature alignment and complicate bi-temporal change detection. To address this issue, we propose FreqHete, a heterogeneous change detection framework that explores both the consistency and disparity between SAR and optical data without auxiliary image generators or adversarial training. First, a style-statistics embedding encoder is developed to inject the deep style statistics of SAR imagery into the low-frequency components of optical features, thereby aligning heterogeneous feature distributions and improving semantic consistency. Second, a frequency-domain interaction decoder is designed to dynamically integrate magnitude and phase information, enabling the model to capture subtle structural variations and enhance sensitivity to change regions. Experiments on the optical-SAR XiongAn benchmark and the satellite-UAV HTCD benchmark show that FreqHete consistently achieves strong overall performance across different heterogeneous change-detection settings while remaining lightweight and efficient. Source code will be released at https://github.com/weiAI1996/FreqHete.