Unsupervised Wavelet Cyclic Translation Network for Change Detection in Heterogeneous Image

Change detection (CD) between heterogeneous images plays a crucial role in monitoring Earth’s dynamic changes. However, the distinct imaging mechanisms of various sensors produce images with differing statistical properties and appearances, making it difficult to directly compare them. To address this issue, we propose an unsupervised CD framework based on wavelet cyclic translation network (WCTNet) for heterogeneous images, which leverages bidirectional image translation and cross-domain difference fusion to highlight changes. First, a pseudo-siamese convolutional neural network is employed to generate an image with similar statistical distribution as the target image, aiming to mitigate modality discrepancies between heterogeneous images. Then, a high-quality difference image is produced by adaptively fusing difference information from both the wavelet and spatial domains, followed by a simple threshold segmentation to identify altered regions. Finally, iterative optimization refines the binary change map to further improve the CD performance. Experimental results on three real datasets with diverse modality combinations and change events demonstrate that the proposed method outperforms several state-of-the-art approaches, confirming its effectiveness and practicability.