A Class-Aware Semi-Supervised Framework for Semantic Segmentation of High-Resolution Remote Sensing Imagery

Advancements in deep learning have significantly improved the semantic segmentation of high-resolution remote sensing imagery (HRSI). However, these approaches typically rely on large-scale, pixel-wise annotations, the acquisition of which is both time-consuming and labor-intensive. This challenge is further amplified in remote sensing, where massive data volumes, strong spatiotemporal heterogeneity, and frequent updates pose additional difficulties. Semi-supervised learning (SSL) has emerged as a promising alternative by utilizing a small amount of labeled data in conjunction with abundant unlabeled data. Despite its potential, the effectiveness of SSL in HRSI is often hindered by two key factors: class imbalance induced by scale variations among objects, and class confusion caused by high intraclass variability and low interclass separability. This study proposes a class-aware semi-supervised semantic segmentation framework that incorporates two key modules: the stratified selection module (SSM) and the category contrastive module (CCM). The SSM adopts a dynamic adaptive thresholding strategy to guide class-specific pseudolabel selection, thus alleviating the impact of class imbalance. The CCM introduces a pixel-to-prototype contrastive mechanism leveraging category-level guidance to improve feature separability and mitigate confusion between similar classes. Experimental results on the ISPRS Vaihingen, Potsdam, and LoveDA datasets validate the effectiveness of the proposed framework under varying annotation ratios. In comparison with state-of-the-art methods, the framework consistently demonstrates superior performance, achieving overall higher accuracy and stability. In particular, it shows robust performance in low-label settings, highlighting its potential for practical applications where annotation resources are limited.