Mars is a dynamic planet exhibiting numerous active surface phenomena, such as recurring slope lineae (RSL), commonly seen on the Martian surface in the mid-latitude regions. Automatic detection of these changes on the Martian surface is pivotal for understanding the evolution and dynamic processes of Mars. Deep-learning models for change detection, such as the Siamese network, have been widely used for identifying changes in images. This paper presents a deep-learning model based on the backbone of the Siamese network, incorporating a spatial attention module and a balanced evaluation method, for detecting dynamic changes on the Martian surface from time-series images. Moreover, we developed a multi-processing web crawler for automatic data retrieval and processing from online image archives, significantly enhancing the efficiency and reach of the change detection method. The effectiveness and reliability of the proposed method have been validated using real time-series Mars images covering typical regions on the Martian surface, focusing on the detection of a typical dynamic phenomenon on Mars, i.e., RSL. The proposed method can automatically retrieve and process data from online image archives, such as the High Resolution Imaging Science Experiment (HiRISE) image archives, and achieves a change detection accuracy of 81.8%. Results indicate that the method can detect subtle changes on the Martian surface from online image archives automatically, showing promising potential for studying the dynamic environment of Mars and enhancing our understanding of Martian surface dynamics.