Answers are grounded in this paper and cite their sources.
The exponential growth of remote sensing image data and the high cost of manual annotation have led to insufficient labeled data, limiting classification performance. Semisupervised methods can address this issue, but most existing approaches lack multi-level consistency constraints in both embedding space and prediction probabilities, resulting in weak feature expressiveness. Additionally, the fine-grained characteristics of remote sensing images, amplified by data augmentation, are often overlooked, leading to overfitting and inaccurate decision boundaries. To address these issues, we propose a multi-level consistency solution that operates at both the embedding space and prediction probability levels. First, we refine the traditional data augmentation strategy and propose a mutual constraint loss function based on this refinement. Second, to address the intra-class variability of remote sensing images, we introduce a novel metric loss function to constrain the embedding space, taking into account both intra-sample consistency and intersample variability. Experimental results on three remote sensing image datasets demonstrate that our proposed method achieves significant performance e improvements compared to existing semi-supervised methods.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: