Semi-Supervised Land-Cover Mapping Based on Multimodal Fusion and Pseudo-Label

Land-cover mapping is of great significance for remote sensing and earth observation. However, due to the high cost of label acquisition, how to use limited labeled samples and multimodal data to achieve large-scale and high-precision land-cover mapping is still a great challenge. In this paper, a multimodal fusion and pseudo-label based method is proposed for semi-supervised land-cover mapping (SLM). For the problem of domain incompatibility, we use strong data enhancement and multimodal fusion module to strengthen the generalization performance of the method from data level and model level respectively. For a large amount of unlabeled data, we combine the pseudo-label self-training technology and propose Fusion-Finetune-Fusion training strategy to achieve large-scale, high-precision land-cover mapping under semi-supervised conditions. In the track SLM of the 2022 Data Fusion Contest (DFC22-SLM), the proposed method achieves a mean intersection over union (mIoU) of 0.4962 in phase 2, ranking fourth place.