Multilabel contrastive learning based remote sensing scene classification via cosine similarity

Multi-Label Classification is a fundamental task in remote sensing, which reflects real-world scenarios by enabling a sample to have more than one label. In the context of multi-class image classification, most recent methods usually make use of the contrastive learning strategy to increase the representative power of the backbone network. Yet, there has been little focus on generalizing the supervised contrastive learning strategy to multi-label classification tasks. In this paper, a new supervised contrastive loss function with a continuous modeling of inter-sample relations is proposed that outperforms common alternative strategies with an optical remote sensing scene classification dataset.