Scholarlink

<p>Accurate cloud segmentation is a critical preprocessing step for terrestrial and atmospheric analysis from satellite imagery. Yet it remains challenging due to spectral ambiguities arising from bright surfaces, such as snow, and from the complex textures of thin clouds. While deep convolutional neural networks (CNNs), particularly UNet-based architectures, have set a strong baseline, they often lack global contextual awareness and struggle with model interpretability [3]. This paper introduces SwinCloudNet, a novel architecture that integrates a Swin Transformer as a powerful encoder to capture long-term dependencies across the image, overcoming the locality constraints of standard convolutions. Furthermore, incorporate an attention-guided fusion mechanism in the decoder to dynamically weight multi-scale features, enhancing edge precision for thin and fragmented clouds. To ensure practical utility, employ explainable AI (XAI) techniques, such as Gradient-weighted Class Activation Mapping (Grad-CAM), to visualise the spectral-spatial features that drive segmentation decisions. On the 38-Cloud and 95-Cloud Landsat-8 datasets, our framework outperforms conventional CNNs, achieving a segmentation accuracy of 97.82% and an IoU of 93.56%. The results underscore the significant benefits of transformer-based global context and attention mechanisms for robust, precise, and interpretable cloud segmentation in multispectral remote sensing.</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

SwinCloudNet: A Transformer-Based Approach for Interpretable Cloud Segmentation in Satellite Imagery

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Thamizharasi Murugaiyan

摘要

Accurate cloud segmentation is a critical preprocessing step for terrestrial and atmospheric analysis from satellite imagery. Yet it remains challenging due to spectral ambiguities arising from bright surfaces, such as snow, and from the complex textures of thin clouds. While deep convolutional neural networks (CNNs), particularly UNet-based architectures, have set a strong baseline, they often lack global contextual awareness and struggle with model interpretability [3]. This paper introduces SwinCloudNet, a novel architecture that integrates a Swin Transformer as a powerful encoder to capture long-term dependencies across the image, overcoming the locality constraints of standard convolutions. Furthermore, incorporate an attention-guided fusion mechanism in the decoder to dynamically weight multi-scale features, enhancing edge precision for thin and fragmented clouds. To ensure practical utility, employ explainable AI (XAI) techniques, such as Gradient-weighted Class Activation Mapping (Grad-CAM), to visualise the spectral-spatial features that drive segmentation decisions. On the 38-Cloud and 95-Cloud Landsat-8 datasets, our framework outperforms conventional CNNs, achieving a segmentation accuracy of 97.82% and an IoU of 93.56%. The results underscore the significant benefits of transformer-based global context and attention mechanisms for robust, precise, and interpretable cloud segmentation in multispectral remote sensing.