TCNet: Temporal Consistency Network for Semisupervised Change Detection

Abstract

Change detection is a challenging task in earth observation. In recent years, deep learning techniques have been widely applied in change detection and achieved impressive progress. However, deep learning based change detection methods heavily rely on a large amount of annotated samples. Labeling for change detection is a time-consuming and labor-intensive task. In order to solve this problem, we propose a novel temporal consistency network (TCNet) for semisupervised change detection. Motivated by the fact that different input sequences have no effect on the prediction results of change detection, our method learns the distribution of unlabeled data by enforce the consistency of the prediction obtained with different input sequences. Specifically, for labeled samples, two segmentation networks with the same structure are trained with two different input sequences. For the unlabeled samples, we perform the forward prediction on the two segmentation networks with corresponding input sequence to obtain two results of change detection. Then, the supervised signals can be generated by minimizing the difference between two predicted results. In this way, the distribution of unlabeled data can be fully explored thus enhancing the generalization of change detection. Experiments on google dataset show the effectiveness of the proposed method.