Abstract: Change detection using synthetic aperture radar (SAR) images based on deep learning has been a significant research topic in the field of remote sensing. However, it is limited by unreliable training samples and highly time-consuming training. Hence, this study proposed a novel unsupervised change detection method using SAR images based on the broad learning system (BLS). First, a reliable pre-classification method is presented by incorporating neighborhood information into similarity operators, adaptive dual-threshold segmentation, superpixel correction, and visual saliency analysis. This pre-classification method generates a pre-classification map and corresponding training samples. Second, the BLS network is trained using the training samples to generate the BLS-based prediction map for change detection. Third, the pre-classification map and the BLS-based prediction map are fused through two-stage voting to generate the final change detection map. The experimental results of five real SAR image datasets show that the proposed method can produce more reliable training samples and achieve higher accuracy in change detection. Moreover, its efficiency is significantly higher than the change detection model using SAR images based on deep learning.