.png){kind=icon}
.png){kind=icon}
Return
Dynamic semantic-collaborative multi-scale semi-supervised segmentation for remote sensing images
.png){kind=icon}
PRE
AI
DOI：10.1016/j.asr.2026.02.022.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Semi-supervised semantic segmentation aims to reduce the reliance on large-scale pixel-level annotations, yet existing methods for remote sensing imagery are often limited by severe scale variations, unreliable pseudo-labels, and insufficient cross-scale feature consistency, leading to degraded performance under low annotation ratios. The objective of this work is to improve segmentation accuracy and annotation efficiency for high-resolution remote sensing images by addressing these limitations. To address these issues, we propose a multi-scale semi-supervised framework with a dynamic semantic synergy mechanism, termed DSCS. The proposed Multi-scale Consistency Learning (MSCL) module employs a dynamic threshold strategy to selectively utilize reliable pseudo-labels and enforce cross-scale feature alignment, while the Bidirectional Symmetric Embedding Learning (BSEL) module further enhances teacher-student collaboration in both feature and output spaces. Experiments on the LoveDA and Potsdam datasets demonstrate that DSCS consistently outperforms state-of-the-art methods under 5 %-20 % labeling settings, especially in complex scenes with pronounced scale diversity. (c) 2026 COSPAR. Published by Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.
Keywords:
Remote Sensing(RS)
Semi-supervised semantic segmentation
Multi-scale
Bidirectional symmetric embedding
Journal
IF:
2.8
Papers:
1.3K
Citations:
2.0W