Introduction
Semantic segmentation in remote sensing images (RSIs) is a critical task for applications ranging from urban planning to environmental monitoring. However, the high cost of pixel-level annotation for satellite and aerial imagery remains a significant bottleneck. Semi-supervised learning (S4) offers a solution by leveraging small amounts of labeled data alongside large unlabeled datasets.
Traditional S4 methods often treat pixels as independent units, which leads to noisy predictions in RSIs where large, coherent structures (like buildings or forests) are common. This paper introduces RegionMatch, a framework that shifts the focus from individual pixels to object-level regions, ensuring consistency and accuracy through pixel-region collaboration.
Motivation: The Reliability of Regions
A core observation driving this research is that while a model may struggle with the classification of an individual pixel, it is often much more confident about the broader "object" or region that pixel belongs to.
As shown in the training performance analysis, region accuracy and regional consistency consistently exceed pixel accuracy. RegionMatch exploits this stability to generate cleaner pseudo-labels for unlabeled data.
Methodology
The RegionMatch framework incorporates three primary innovations to bridge the gap between pixel-level detail and region-level stability:
- Object Region Classification: The model groups pixels into regions using a connected component algorithm. A dedicated Region Classifier then identifies these objects, providing a high-level structural view of the scene.
- Pixel-Region Synergy Pseudo-Labeling (PRSPL): This consensus strategy only generates pseudo-labels when the pixel-level and region-level predictions agree. By using a dynamic entropy threshold, the model initially learns from high-certainty areas and gradually tackles more complex structures.
- Region-Structure Aware Correlation Consistency (RSCC): This module ensures consistency across two scales:
- Cross-Image (CIR²R): Aligns feature signatures of similar regions (e.g., forests) across different images.
- Intra-Region (PRCC): Forces pixels within a region to align with the overall region feature, ensuring internal consistency without the high computational cost of pixel-to-pixel comparisons.
Experimental Results
RegionMatch was evaluated on three major remote sensing benchmarks: Vaihingen, Potsdam, and LoveDA. It consistently outperformed state-of-the-art methods like FixMatch and UniMatch across various labeling ratios.
- Vaihingen: Achieved 69.25% mIoU with only 1/8 labeled data, an 8% improvement over the supervised baseline.
- Potsdam: Reached 80.24% mIoU at the 1/8 ratio, surpassing the next best method by over 1%.
- LoveDA: Demonstrated robustness in complex urban/rural scenes, achieving 52.55% mIoU at a 1/16 labeling ratio.
Qualitative analysis via t-SNE plots confirmed that RegionMatch produces tighter and more distinct clusters for land-cover classes compared to standard baselines, indicating superior feature representation.
Conclusion
RegionMatch demonstrates that leveraging region-level information is key to effective semi-supervised semantic segmentation in remote sensing. By enforcing collaboration between pixel and region classifiers, the framework generates high-quality pseudo-labels and achieves state-of-the-art performance with minimal supervision. Future work may explore more sophisticated unsupervised region definition methods, such as superpixel algorithms, to further enhance feature stability.
Summary updated