Change detection of surface water in remote sensing images based on fully convolutional network

This study presents a new approach based on fully convolutional networks (FCN) to detect changes in surface water. The proposed method can be divided into three steps: (1) training the FCN using color-infrared (CIR) images from the Coastwide Reference Monitoring System (CRMS) dataset with two classes, such as water and land; (2) passing the multitemporal images respectively through the pre-trained FCN and generating a difference image (DI) from score maps of the last prediction layers; and (3) determining optimal threshold values using fuzzy entropy and discriminating between changed and unchanged pixels in the DI. This method has the advantage of effectively learning the spatial and spectral characteristics of water bodies from large remote-sensing datasets, and it would be helpful to analyze and monitor changes in newly obtained images without ground truth. The experimental results obtained using the multitemporal CRMS data demonstrated the effectiveness of this deep-learning approach for detecting changes in remote-sensing images, as compared other traditional methods for change detection.