Scholarlink

<p>Ground-based cloud automatic observation is crucial for meteorological research and disaster early warning, providing essential data for weather prediction and environmental monitoring. However, the scarcity of annotated data significantly limits the performance of deep learning-based semantic segmentation models, which are vital for accurate cloud image analysis. To address this challenge, we propose a novel self-training framework for semantic segmentation of Ground-Based Cloud Images (GBCIs). Our method leverages easily accessible unlabeled images and enhances model generalization by injecting strong noise during the training process. Specifically, we explore four strong noise strategies—Blur, Erasing, Color Jitter, and Grayscale—applied to pseudo labels generated during the self-training process. Experimental results show that Blur and Erasing strategies significantly improve segmentation performance across different network architectures, such as UNet, PSPNet, and DeepLabv3+. In contrast, Color Jitter and Grayscale augmentations degrade accuracy by disrupting critical color information necessary for distinguishing subtle cloud features. This finding emphasizes the importance of carefully tailored augmentation strategies. Our comprehensive evaluation demonstrates that, when appropriately selected, noise injection can effectively utilize vast amounts of unlabeled data, enhancing segmentation accuracy and robustness. The proposed self-training framework shows great promise for GBCI segmentation, providing a scalable and efficient solution when annotated data resources are limited.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Applying self-training for semantic segmentation of ground-based cloud images

  • 期刊论文

  • 发表日期：

    2026年5月15日

查看全文
  • Jifei Sun,
  • Kebin Jia

摘要

Ground-based cloud automatic observation is crucial for meteorological research and disaster early warning, providing essential data for weather prediction and environmental monitoring. However, the scarcity of annotated data significantly limits the performance of deep learning-based semantic segmentation models, which are vital for accurate cloud image analysis. To address this challenge, we propose a novel self-training framework for semantic segmentation of Ground-Based Cloud Images (GBCIs). Our method leverages easily accessible unlabeled images and enhances model generalization by injecting strong noise during the training process. Specifically, we explore four strong noise strategies—Blur, Erasing, Color Jitter, and Grayscale—applied to pseudo labels generated during the self-training process. Experimental results show that Blur and Erasing strategies significantly improve segmentation performance across different network architectures, such as UNet, PSPNet, and DeepLabv3+. In contrast, Color Jitter and Grayscale augmentations degrade accuracy by disrupting critical color information necessary for distinguishing subtle cloud features. This finding emphasizes the importance of carefully tailored augmentation strategies. Our comprehensive evaluation demonstrates that, when appropriately selected, noise injection can effectively utilize vast amounts of unlabeled data, enhancing segmentation accuracy and robustness. The proposed self-training framework shows great promise for GBCI segmentation, providing a scalable and efficient solution when annotated data resources are limited.