Unique Perturbation Methods Exploitation for Semi-Supervised Remote Sensing Image Semantic Segmentation

Abstract

Deep learning has significantly improved the accuracy of remote sensing semantic segmentation, yet its effectiveness is often constrained by the limited availability of annotated training samples. Semi-supervised learning (SSL) addresses this challenge by utilizing abundant unlabeled data, reducing dependence on manual annotations. However, current consistency regularization-based SSL methods, primarily developed for natural images, struggle to produce adequate perturbation diversity for robust model training in remote sensing image segmentation. In this work, we propose FusionMatch, a novel SSL framework featuring two perturbation mechanisms - NIRPerb and PSPerb - specifically designed for remote sensing imagery. NIRPerb utilizes near-infrared spectral data to enhance perturbation diversity. PSPerb adopts differentiated pan-sharpening fusion strategies to expand the perturbation space. Extensive experiments on both a building extraction dataset and a multi-class dataset demonstrate that FusionMatch outperforms state-of-the-art SSL methods in segmentation accuracy and robustness.