Multi-task Learning Based Single Temporal Remote Sensing Image Change Detection.

Peixuan Yang,Yichao Cao,Xiaobo Lu

DOI: https://doi.org/10.1145/3633637.3633670

2023-01-01

Abstract:Change Detection is a classic and fundamental task in remote sensing(RS), and has accomplish remarkable achievement on bitemporal image based methods. However, bitemporal images are very expensive and needs professional researchers to align and label the paired images. Furthermore, most bitemporal image based methods are only semantic-aware not instance-aware, which indicates these methods only use paired images’ semantic information. Thus, we propose a method that only needs unpaired RS images with semantic labels. In this paper, we introduce multi-task learning into change detection progress, which comprises four major parts: Boundary-Aware Module, Contextual-Aware Module, Segmentation Module, Change Detection Module. At first, we construct Boundary-Aware Module and Contextual-Aware Module to extract sharp boundary and dense contextual features. Then we feed the feature maps to Change Detection Module and Segmentation Module to accomplish the subtasks. The comprehensive experiments on a single temporal image dataset demonstrate the efficiency of our approach over other bitemporal image based methods, and the experimental results shows our approach is competitive with other methods while all methods was trained on bitemporal image dataset.