A deep supervised network (DSNet) for remote sensing image change detection (CD) was built with the Unet++ to address the inadequate feature extraction and poor ability of remote sensing image CD in most fullly convolutional networks. A multi-scale residual module was designed to replace the conventional convolution layer. The problem of single feature extraction was resolved by combining the spatial and spectral properties of twin networks to obtain the semantic information of remote sensing images at various levels. The horizontal output layer of model was built to implement the deep supervision process of node feature aggregation from shallow to deep features at the decoding end of the model. The fusion results of different features with information differentiation were transmitted to the normalization-based attention module (NAM). The information weight of the changed region was enhanced without introducing additional parameters. Experimental results showed that the recall rate and accuracy of the proposed model in the remote sensing image CD task were 90.39% and 92.04%, respectively, and the parameter quantity and calculation quantity of the model were 6.38 M and 60 G. Comparise with different network models indicates that the proposed method has the advantages of high detection accuracy, fast speed and lightweight.