RSMatch: Semi-supervised Learning with Adaptive Category-Related Pseudo Labeling for Remote Sensing Scene Classification

Remote sensing scene classification (RSSC) has become a hot and challenging research topic in recent years due to its wide applications. Due to the development of convolutional neural networks (CNN), the data-driven CNN-based methods have achieved expressive performance in RSSC. However, the lack of labeled remote sensing scene images in real applications make it difficult to further improve their performance of classification. To address this issue, we propose a novel adaptive category-related pseudo labeling (ACPL) strategy for semi-supervised scene classification. Specifically, ACPL flexibly adjusts thresholds for different classes at each time step to let pass informative unlabeled data and their pseudo labels according to the model’s learning status. Meanwhile, our proposed ACPL dose not introduce additional parameters or computation. We apply ACPL to FixMatch and construct our model RSMatch. Experimental results on UCM data set have indicated that our proposed semi-supervised method RSMatch is superior to its several counterparts for RSSC.