Remote sensing image has the characteristics of real-time, periodicity and wide monitoring range. It can quickly and accurately obtain water area, distribution and other information, which is of great significance to the utilization and development of water resources, agricultural irrigation, flood disaster assessment and so on. Since traditional water information extraction methods only use part of image band information, the accuracy of water information extraction is low and has certain limitations. In recent years, convolutional neural network technology has developed rapidly and achieved good results in water information extraction from remote sensing images. Therefore, in this paper, Pyramid Scene Parsing Neural Network (PSPNet) was used to extract water information from Ziyuan-3 multispectral remote sensing images, to make sample sets of water, and train the convolutional neural network model. Compared with the traditional normalized difference water index (NDWI) and support vector machine (SVM), the results show that PSPNet has the highest accuracy and the lowest misclassification rate.