Change detection (CD) in remote sensing images has long been of great interest to researchers. Due to the substantial time and effort required for data annotation in remote sensing, semi-supervised methods have attracted extensive attention as a promising solution for achieving satisfactory performance under limited samples. However, existing semi-supervised approaches often encounter significant challenges, most notably class imbalance and subtle changes in feature distributions, making it difficult to distinguish changes from unchanged regions. Here, we propose a novel strategy, named NF-SemiCD, which incorporates normalizing flows in a semi-supervised architecture. We leverage normalizing flows to characterize the feature distribution of unchanged regions and derive a probability distribution model. Its sensitivity to changes in probability allows us to get a probability feature map, which provides useful information on deep difference features. To this end, we devise a three-stage training scheme: (1) training an encoder-decoder network with labeled data, (2) training a normalizing flow decoder on labeled data, and (3) training the encoder-decoder network with all data. Experiments on three benchmark datasets demonstrate that NF-SemiCD outperforms existing state-of-the-art methods, highlighting its potential for improved change detection under semi-supervised settings.
Yu et al. (Fri,) studied this question.