Semi-Supervised Remote Sensing Segmentation via Foundation Model Distillation and Consistency Learning
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Semantic segmentation of remote sensing imagery is fundamental to urban planning, environmental monitoring, and ecosystem management, yet its progress is constrained by the high cost of pixel-level annotations. Semi-supervised learning alleviates this issue by exploiting unlabeled data, but existing methods often suffer from pseudo-label noise and insufficient exploitation of structural information when annotations are scarce. To address these limitations, we propose FDMatch, a semi-supervised segmentation framework that integrates foundation model distillation with multi-level consistency learning. A pretrained remote sensing foundation model is employed as a knowledge source, from which rich spatial priors and high-level semantics are adaptively distilled into a lightweight segmentation network, enhancing feature discrimination under limited supervision. Moreover, a multi-level consistency regularization strategy is introduced to overcome the limitation of conventional pixel-wise constraints by modeling inter-class relationships and inter-pixel correlations, thereby providing more reliable supervisory signals from unlabeled data. Extensive experiments on two benchmark datasets demonstrate that FDMatch consistently outperforms existing semi-supervised methods across different label ratios, particularly under extremely limited annotation scenarios.
Fields of Study & Subjects
Remote-Sensing Image Classification
Advanced Neural Network Applications
Domain Adaptation and Few-Shot Learning
Publication details
- Journal / source
- IEEE International Geoscience and Remote Sensing Symposium (IGARSS)
- Pages
- 5270–5274
- Publication date
- 1 August 2026
- Language
- EN