Few-shot segmentation of remote sensing imagery with textual prompts

Remote sensing image segmentation is crucial for Earth observation and scene understanding, yet its performance is often hindered by the scarcity of labelled data. Although few-shot segmentation presents a promising solution, most existing methods are designed for natural images and exhibit poor performance on remote sensing data due to domain shifts and spectral ambiguity. To address these challenges, we propose a novel text-prompted few-shot segmentation model that integrates contrastive learning and graph neural networks. Textual information is introduced to enrich image feature representations, enhancing their discriminability, while a graph structure models inter-region relationships and propagates label information to increase the number of labelled samples. These two strategies work in synergy to enhance segmentation performance. Experiments on the WHDLD and GID5 datasets show that the proposed model substantially outperforms existing few-shot segmentation methods in terms of mean Intersection over Union (mIoU) and overall accuracy. These results confirm the model’s effectiveness and strong generalization capabilities, particularly in scenarios with limited labelled data and complex spectral characteristics.