Learning Positive–Negative Prompts for Open-Set Remote Sensing Scene Classification

Most remote sensing scene classification (RSSC) methods are primarily built upon the closed-set assumption, which assumes that all test samples definitely belong to one of the categories seen during training. However, practical applications are usually open environments, where samples of other categories that have never been seen during training will appear, which is called open-set RSSC (OS-RSSC). These methods may mistakenly classify samples of unseen categories into those seen categories, resulting in a decrease in application potential. In this article, we propose a positive–negative prompt learning (PNPL) framework for OS-RSSC. PNPL aims to tune the powerful contrastive language-image pretraining (CLIP) model for OS-RSSC through learning positive and negative prompts. First, positive textual prompts and visual prompts are trained to provide the model with basic classification capabilities for known classes. Then, negative textual prompts are indirectly learned from positive prompts and images, enabling the model to capture the semantics of unknown classes. PNPL significantly increases the discriminative power between known and unknown classes, enhancing the model’s ability to accurately distinguish them. Extensive experiments on three RSSI datasets have shown that PNPL outperforms compared methods. Code is available at https://github.com/ChenHanlizi/PNPL.