Multi-Stage Pseudo-Label Iteration Framework for Semi-Supervised Land-Cover Mapping

Land-cover mapping is a pivotal pathway for Earth observation. Nevertheless, the lack of labeled data and the domain gap of different mapping regions are still challenges inhibiting the large-scale implementation of common mapping methods. In this article, a multi-stage pseudo-label iteration framework is presented for the semi-supervised land-cover mapping track of the 2022 Data Fusion Contest (DFC-SLM). The proposed framework combines the multi-stage training process with the pseudo-label technique to tackle the issues of large-scale land-cover mapping task when limited labeled samples are available. Firstly, the multi-stage training process promotes to sufficiently explore the discriminative and robust features of land covers with limited labeled data, where the proportion of samples in confusing classes gradually increases. Secondly, the pseudo-label technique generates extra supervision information from the unlabeled imagery. Overall, experimental results obtained from several cities in France achieved a mIoU of 52.96%, and won 2nd place on the final leaderboard of the 2022 DFC-SLM.