Scholarlink

<p>Accurate Green View Index (GVI) estimation is crucial for ecological assessment in complex landscapes. Jiangnan classical gardens and water towns pose unique challenges including high-contrast facades, water reflections, and fine-grained occlusions. This study proposes a precision-first segmentation framework based on ResNet50+Convolutional Block Attention Module (CBAM)+DeepLabV3, integrating Atrous Spatial Pyramid Pooling (ASPP), attention-based reweighting, and adaptive thresholding. Trained on 1,013 annotated images, the model achieved Intersection over Union (IoU) of 0.699, Precision of 0.806, and F1-score of 0.823, outperforming traditional methods (+41.2% IoU) and deep learning baselines (+7.2% IoU). Error analysis confirmed a false positive rate of 6.1%. An interactive Streamlit platform enables real-time deployment, advancing vegetation segmentation methodology for heritage conservation and ecological monitoring.</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

Precision-focused green view index estimation in Jiangnan classical gardens and water towns using attention-enhanced deep learning segmentation

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Xiao Zheng,
  • You Lv

摘要

Accurate Green View Index (GVI) estimation is crucial for ecological assessment in complex landscapes. Jiangnan classical gardens and water towns pose unique challenges including high-contrast facades, water reflections, and fine-grained occlusions. This study proposes a precision-first segmentation framework based on ResNet50+Convolutional Block Attention Module (CBAM)+DeepLabV3, integrating Atrous Spatial Pyramid Pooling (ASPP), attention-based reweighting, and adaptive thresholding. Trained on 1,013 annotated images, the model achieved Intersection over Union (IoU) of 0.699, Precision of 0.806, and F1-score of 0.823, outperforming traditional methods (+41.2% IoU) and deep learning baselines (+7.2% IoU). Error analysis confirmed a false positive rate of 6.1%. An interactive Streamlit platform enables real-time deployment, advancing vegetation segmentation methodology for heritage conservation and ecological monitoring.