To achieve efficient change detection in remote sensing images, a lightweight Siamese network for change detection (LS-CDNet) was proposed. LS-CDNet is constructed based on a Siamese network architecture, with the lightweight network MobileNetV2 used as the backbone. A cascaded attention module is designed to optimize the low-level features and enhance the boundary information of the regions. In order to consider both the convergence efficiency of model training and the learning ability of imbalanced datasets for positive and negative samples, a weighted combination of the BCE Loss and Dice Loss is used to optimize the model learning strategy. Experimental results on the LEVIR-CD and CDD datasets demonstrate that LS-CDNet achieves precision rates of 88.06% and 90.12% respectively, with a model parameter size of 3.76 M and a computational cost of 2.18 G (FLOPs). The performance of LS-CD-Net outperforms other comparative methods.