Scholarlink

<p>Understanding the changing landscapes of urban and rural areas is essential for sustainable planning and informed decision making, a process fundamentally driven by Land Use and Land Cover (LULC) analysis. LULC analysis using satellite images offers accurate, large-scale and timely monitoring of land surface changes. To provide a precise, pixel-level representation of land cover boundaries, segmentation approaches are preferred over classification approaches. Existing segmentation approaches often fail to capture essential local and global spatial-contextual features, leading to reduced accuracy in LULC analysis. Additionally, the complexity of satellite imagery makes these methods computationally intensive and less efficient. Hence, this paper proposes a novel satellite image segmentation approach that combines U-Net with the MobileNet backbone deep learning technique, along with position and channel attention modules at the bottleneck for effective segmentation. U-Net has expanding and contracting paths that enable precise segmentation. MobileNet is used as a lightweight encoder structure because it uses depthwise separable convolutions and it is trained on ImageNet pretrained weights, which enhances the feature extraction process. In the proposed U-Net model, position attention captures spatial relationships within the image, whereas channel attention is used to employ a differentiable between two classes. This approach can learn the hierarchical features by itself and integrates global and local context features, enabling comprehensive image analysis. The proposed model has been trained on the Gaofen -2 satellite images (GID) dataset. This dataset has 15 classes, denoting land use and land cover patterns. The proposed model gives an accuracy of 79.05% with an IoU score of 0.5549 proving to be a robust model for image segmentation. Compared to the base model (PT-GID), which achieves an accuracy of 70.04%, our study demonstrates a notable improvement with an accuracy of 9.01% which is due to the attention blocks that are added to U-Net, which considers global and local context.</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

Enhancing satellite image segmentation for land use and land cover analysis using attention-based deep learning approach

  • 期刊论文

  • 发表日期：

    2026年1月17日

查看全文
  • K.S Gayathri,
  • R Ritheesh Kumar,
  • Varshini Ramesh,
  • Surendar Natarajan

摘要

Understanding the changing landscapes of urban and rural areas is essential for sustainable planning and informed decision making, a process fundamentally driven by Land Use and Land Cover (LULC) analysis. LULC analysis using satellite images offers accurate, large-scale and timely monitoring of land surface changes. To provide a precise, pixel-level representation of land cover boundaries, segmentation approaches are preferred over classification approaches. Existing segmentation approaches often fail to capture essential local and global spatial-contextual features, leading to reduced accuracy in LULC analysis. Additionally, the complexity of satellite imagery makes these methods computationally intensive and less efficient. Hence, this paper proposes a novel satellite image segmentation approach that combines U-Net with the MobileNet backbone deep learning technique, along with position and channel attention modules at the bottleneck for effective segmentation. U-Net has expanding and contracting paths that enable precise segmentation. MobileNet is used as a lightweight encoder structure because it uses depthwise separable convolutions and it is trained on ImageNet pretrained weights, which enhances the feature extraction process. In the proposed U-Net model, position attention captures spatial relationships within the image, whereas channel attention is used to employ a differentiable between two classes. This approach can learn the hierarchical features by itself and integrates global and local context features, enabling comprehensive image analysis. The proposed model has been trained on the Gaofen -2 satellite images (GID) dataset. This dataset has 15 classes, denoting land use and land cover patterns. The proposed model gives an accuracy of 79.05% with an IoU score of 0.5549 proving to be a robust model for image segmentation. Compared to the base model (PT-GID), which achieves an accuracy of 70.04%, our study demonstrates a notable improvement with an accuracy of 9.01% which is due to the attention blocks that are added to U-Net, which considers global and local context.