Objectives Recently, it is gradually becoming important to timely monitor the accurate distribution and change of water resource on the earth due to the shortage of water resource globally. Water extraction from remotely sensed image not only is able to provide data support for water ecosystem analysis and risk prediction, but also can alleviate the currently serious ecosystem environment and contribute to sustainable development of water resource and environment protection. Because of its powerful feature representation, deep learning has an important influence on computer vision and image recognition domains. Fully convolutional networks (FCN) is the first end-to-end fully convolutional network for image pixel-wise prediction and starts the image semantic segmentation. Hereafter, other variants of image semantic segmentation based on FCN are also put up. However, the border region between water and land is still misclassified during water extracting based on FCN or its variants. Therefore, we put up water extraction of high spatial resolution remotely sensed image fusing scale attention mechanism based on squeeze-and-excitation weighted FCN (SE-wFCN).

Methods We utilize GF-1 remotely sensed image to do some experiments by comparing with FCN and its variants. The proposed method has the following contribution of three aspects: (1) In order to take advantage of attention mechanism, this method applies it to features of different scales and then fuses those features more reasonably and more scientifically to reduce the possibility of misclassification of water pixels as far as possible. (2) Due to the model complexity of FCN and its possibility of overfitting, we add one auxiliary classifier module to FCN model as one branch during the training phase, which is able to help to reduce the possibility of overfitting. (3) Considering that the border pixel's position between water and land is very important to water extraction and the position information is actually ignored when computing the loss during training phase, we pay more attention to the border pixel by putting more weight in the loss function.

Results Experimental analysis shows that the proposed method is feasible and effective.

Conclusions It can achieve the more satisfying of water extraction than these current popular deep learning models (i.e. DeeplLabV3, U-Net, U2Net, MECNet, MSResNet and SADA-Net), especially in the border between water and land.