Negative Samples Mining Matters: Reconsidering Hyperspectral Image Classification With Contrastive Learning

In recent years, there have been significant advancements in hyperspectral image (HSI) classification methods using contrastive learning. However, these methods often fail to effectively screen and mine negative samples during the construction of contrastive learning pairs. This oversight introduces negative samples that belong to the same class as the positive samples, thereby limiting the effectiveness of contrastive learning. To address this issue, we propose a novel HSI classification framework based on contrastive learning that flexibly supports the selection and mining of negative samples. Specifically, before training the contrastive learning task, we use pseudolabel information to guide the mining of negative samples, eliminating those with pseudolabels of the same class as the anchor samples. This approach strengthens the alignment between the optimization directions of the contrastive learning task and the classification task. In addition, we improve the contrastive learning process by introducing a controlled mixture of hard and easy negative samples, which enhances the accuracy of HSI classification. The pivotal characteristic of this study lies in enhancing the effectiveness of HSI classification based on contrastive learning through effective filtering and mining of negative samples. Experimental comparisons across multiple public datasets demonstrate the superiority of our proposed method over state-of-the-art algorithms.