<p>Change detection (CD) is essential for remote sensing applications. Supervised learning models, particularly supervised deep learning, rely heavily on large amounts of high-quality samples. In contrast, unsupervised learning models mainly utilize pixel-level information, which makes it difficult to capture image object-level features. These challenges significantly limit the application of CD technology. To overcome these challenges, we propose an unsupervised deep learning CD method combining the segment anything model (SAM) with multivariate alteration detection (MAD). The proposed method uses two-phase images with a unified color style to calculate true color and grayscale difference images, and segments the true color difference image using SAM. MAD-expectation maximization is then employed to generate object-level difference images from the segmented true color images. By fusing the object-level true color difference image and the grayscale difference image, a change probability image is generated. A dynamic CD threshold is set to achieve CD result from the probability image. The results show that compared to other methods such as MAD and DSFANet, the proposed method improves the Kappa coefficient by at least 12%, achieving an overall accuracy above 0.94. This method is versatile across scenarios and data types, offering a more robust solution for intelligent remote sensing image interpretation.</p>