Time-series PolSAR are capable for continuous change monitoring of natural resources and urban land-covers regardless of weather and lighting conditions. However, in the big SAR data era, the scarcity of labelled PolSAR samples poses new challenge to the traditional change detection methods. To reduce the dependence on labeled samples and ensure the efficiency of long time-series PolSAR interpretation, an unsupervised and pseudo label-based change detection method is proposed. First, the similarity maps of time-series PolSAR are gauged by three selected polarimetric distance measurements (PDM), which are suitable for PolSAR distribution characteristics and have the potential to reflect PolSAR changes. Second, the high-confidence changed pseudo-samples are selected based on the similarity maps, and the unchanged pseudo-samples are selected based on the non-similarity maps. Thirdly, the limited selected pseudo-samples (changed and unchanged) and multi-dimensional features are used to train the ConvLSTM network for change detection, and the input features include the T₃ coherence matrix elements of time-series PolSAR and the aforementioned PDMs. Finally, the change detection results based on pseudo-samples and the ConvLSTM network can be obtained, without additional manual labels. Adequate experiments are conducted on Radarsat-2, UAVSAR full-polarized, and Sentinel-1 dual-polarized datasets, achieving improved unsupervised change detection accuracy at 89.59–93.24%.

Limited labels and detailed changed land-cover interpretation requirements pose challenges for time-series PolSAR change monitoring research. Accurate labels and supervised models are difficult to reuse between massive unlabeled time-series PolSAR data due to the complex distribution shifts caused by different imaging parameters, scene changes, and random noises. Moreover, many related methods can only detect binary changes in PolSAR images and struggle to track the detailed land cover changes. In this study, an unsupervised cross-domain method based on limited-label transfer learning and a vision transformer (LLTL-ViT) is proposed for PolSAR land-cover change monitoring, which effectively alleviates the problem of difficult label reuse caused by domain shift in time-series SAR data, significantly improves the efficiency of label reuse, and provides a new paradigm for the transfer learning of time-series polarimetric SAR. Firstly, based on the polarimetric scattering characteristics and manifold-embedded distribution alignment transfer learning, LLTL-ViT transfers the limited labeled samples of source-domain PolSAR data to unlabeled target-domain PolSAR time-series for initial classification. Secondly, the accurate samples of target domains are further selected based on the initial transfer classification results, and the deep learning network ViT is applied to classify the time-series PolSAR images accurately. Thirdly, with the reliable secondary classification results of time-series PolSAR images, the detailed changes in land cover can be accurately tracked. Four groups of cross-domain change monitoring experiments were conducted on the Radarsat-2, Sentinel-1, and UAVSAR datasets, with about 10% labeled samples from the source-domain PolSAR. LLTL-ViT can reuse the samples between unlabeled target-domain time-series and leads to a change detection accuracy and specific land-cover change tracking accuracy of 85.22–96.36% and 72.18–88.06%, respectively.