We propose an approach for multiple change detection in multispectral remote sensing images based on joint sparse representation. The principal idea is that each change class lies in a low-dimensional space, in which the change vectors can be represented by a linear combination of a few representation atoms. Our method includes two stages: (1) in the learning stage, we learn a subdictionary for each change class from the training samples; and (2) in the reference stage, each change pixel vector is represented with respect to all subdictionaries and assigned to the class with minimum representation errors. Furthermore, taking into account the spatial contextual information, we propose a joint sparsity model to represent each pixel and its similar neighbors simultaneously, thereby enhancing the robustness of the representation to noise. To validate the effectiveness of the proposed method, we choose Shenzhen, China, as the study area in the context of fast urban growth. During the experiments, the proposed method achieves better results on two Landsat Enhanced Thematic Mapper Plus images than does another state-of-the-art supervised change-detection method.