A Self-Reinforcing Prototype Framework to Mitigate Pseudo-label Degradation in Semi-Supervised Remote Sensing Segmentation

Semi-supervised semantic segmentation (SSSS) has emerged as an effective strategy to alleviate the dependence on costly pixel-level annotations in remote sensing imagery by exploiting limited labeled data alongside abundant unlabeled samples. However, conventional confidence-based pseudo-labeling often produces noisy and semantically inconsistent supervision, degrading feature representations and limiting overall performance. To address these challenges, this paper proposes a Self-Reinforcing Prototype Framework (SRPF) that introduces class-level prototypes to enhance pseudo-label generation and improve semantic consistency. The proposed framework consists of two key components: Prototype-enhanced Pseudo-label Generation (PPG), which produces pseudo-labels by matching feature embeddings with dynamic class-specific prototypes, and Adaptive Prototype Initialization (API), which computes prototypes from labeled feature centroids to provide stable and discriminative starting points. Prototypes are further refined through momentum-based updates guided by high-confidence predictions, enabling continuous adaptation. The mutual reinforcement between PPG and API establishes a self-improving cycle in which more accurate prototypes yield better pseudo-labels, and enhanced pseudo-labels, in turn, refine prototype representations. Extensive experiments on five benchmark datasets (GID-15, MER, MSL, Vaihingen, and DFC22) validate the effectiveness of SRPF, achieving consistent performance improvements, with ablation studies confirming mIoU gains of 1.42% and 1.26% contributed by PPG and API, respectively.