Multi-class change detection using a multi-task Siamese network of remote sensing images

The accurate acquisition of land cover/use changes and their types is critical to territorial space planning, ecological environment monitoring, and disaster assessment. However, most current studies on the change detection focus on binary change detection. This study proposed a multi-class change detection method using a multi-task Siamese network of remote sensing images. First, an object-oriented unsupervised change detection method was employed to select areas that were most/least prone to change in the new and old temporal images. These areas were used as samples for the multi-task Siamese network. Subsequently, the multi-task Siamese network model was used to learn and predict the new and old temporal land-use maps and binary change maps. Finally, the final multi-class change detection results were derived from these maps. The multi-task Siamese network was tested based on the images from the Third National Land Survey and corresponding land-use maps. The results demonstrate that the method proposed in this study is applicable to the change detection cases where changed and unchanged samples lack but there are available historical thematic maps.