Scholarlink

<p>Understanding land cover is crucial for a wide range of geospatial applications, from managing natural resources to monitoring environmental changes and classifying satellite imagery. While Convolutional Neural Networks (CNNs) have been the go-to method for these tasks due to their strength in capturing local image features, they often fall short when it comes to modeling long-range dependencies and spatial relationships. The success of transformer models depends on large-scale datasets, yet satellite imagery datasets typically have small sample sizes. To solve these issues, we introduce an innovative approach that combines the benefits of Swin Transformers and Graph Neural Networks (GNNs) to improve Land Use and Land Cover (LULC) classification. The former are effective at capturing both local and global features through a hierarchical self-attention mechanism, while the latter help model spatial and contextual relationships across different regions in the image. Aiming to build an innovative model, we enhance Swin-GNNet with a fuzzy block to handle data uncertainty and improve model reliability. Experimental results on the riverscape dataset reveal that the fuzzified Swin-GNNet shows superior performance than other deep neural network models, achieving over 87% accuracy in distinguishing all LULC classes such as water, urban, rural, agriculture, forest, barren, and wetland.</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

Swin Transformer–Graph Neural Network (Swin–GNNet) for Land Use Land cover classification: a fuzzy hybrid approach

  • 期刊论文

  • 发表日期：

    2026年6月15日

查看全文
  • V. Anitha,
  • S. Kalaiselvi,
  • R. Inisha Sallove

摘要

Understanding land cover is crucial for a wide range of geospatial applications, from managing natural resources to monitoring environmental changes and classifying satellite imagery. While Convolutional Neural Networks (CNNs) have been the go-to method for these tasks due to their strength in capturing local image features, they often fall short when it comes to modeling long-range dependencies and spatial relationships. The success of transformer models depends on large-scale datasets, yet satellite imagery datasets typically have small sample sizes. To solve these issues, we introduce an innovative approach that combines the benefits of Swin Transformers and Graph Neural Networks (GNNs) to improve Land Use and Land Cover (LULC) classification. The former are effective at capturing both local and global features through a hierarchical self-attention mechanism, while the latter help model spatial and contextual relationships across different regions in the image. Aiming to build an innovative model, we enhance Swin-GNNet with a fuzzy block to handle data uncertainty and improve model reliability. Experimental results on the riverscape dataset reveal that the fuzzified Swin-GNNet shows superior performance than other deep neural network models, achieving over 87% accuracy in distinguishing all LULC classes such as water, urban, rural, agriculture, forest, barren, and wetland.