In order to study the potential of water extraction from multispectral camera (PMS) images of GF-6, four kinds of neural networks, including full convolutional neural network (FCN-8s), U-Net and U-Net optimization (VGGUnet1, VGGUnet2), were constructed for water extraction studies. Based on the water extraction results, the best model was determined as VGGUnet1; then a VGGUnet1 network model based on the combined loss function Focal-Dice-Water loss (FD-Water loss) was proposed. Compared with the normalized water index (NDWI) threshold method, maximum likelihood classification method, and support vector machine classification method, the results show that the VGGUnet1 network model based on the FD-Water loss function can effectively extract the water body target, enhance the recognition ability of the water body in a small area, reduce the phenomenon of misdivision and leakage of the water body, and improve the extraction effect of the water body. It is concluded that the full convolutional neural network is feasible in water extraction of GF-6 remote sensing images, which provides a reference for further research and application in this field.