Water body extraction based on SAR images is widely used in flood monitoring and other fields. The water body extraction method based on threshold segmentation is easy to classify mountain shadows as water bodies for wrong extraction. Traditional machine learning classification methods require manual extraction of effective features, which is inefficient and time-consuming. This paper proposes a SAR image water extraction method based on convolutional neural network. First, the SAR image is divided into blocks, and the SAR image features are automatically learned through multi-layer convolution and pooling operations based on image blocks. Finally, the Sigmoid classifier is used to classify the pixels into water and non-water to realize the extraction of water body. Experiments based on SAR data obtained by Sentinel-1A verify the effectiveness of this method. The recall rate and accuracy rate of water extraction can reach 99%, and the performance is better than the OTSU threshold method and the SVM method based on texture features. This method overcomes the influence of mountain shadows on water body extraction, and its ability to automatically learn features can achieve efficient water body extraction.