Abstract
Remote Sensing Change Detection (RS-CD) aims to detect relevant changes fromMulti-Temporal Remote Sensing Images (MT-RSIs), which aids in various RSapplications such as land cover, land use, human development analysis, anddisaster response. The performance of existing RS-CD methods is attributed totraining on large annotated datasets. Furthermore, most of these models areless transferable in the sense that the trained model often performs verypoorly when there is a domain gap between training and test datasets. Thispaper proposes an unsupervised CD method based on deep metric learning that candeal with both of these issues. Given an MT-RSI, the proposed method generatescorresponding change probability map by iteratively optimizing an unsupervisedCD loss without training it on a large dataset. Our unsupervised CD methodconsists of two interconnected deep networks, namely Deep-Change ProbabilityGenerator (D-CPG) and Deep-Feature Extractor (D-FE). The D-CPG is designed topredict change and no change probability maps for a given MT-RSI, while D-FE isused to extract deep features of MT-RSI that will be further used in theproposed unsupervised CD loss. We use transfer learning capability toinitialize the parameters of D-FE. We iteratively optimize the parameters ofD-CPG and D-FE for a given MT-RSI by minimizing the proposed unsupervised``similarity-dissimilarity loss''. This loss is motivated by the principle ofmetric learning where we simultaneously maximize the distance between changepair-wise pixels while minimizing the distance between no-change pair-wisepixels in bi-temporal image domain and their deep feature domain. Theexperiments conducted on three CD datasets show that our unsupervised CD methodachieves significant improvements over the state-of-the-art supervised andunsupervised CD methods. Code available at https://github.com/wgcban/Metric-CD