Scholarlink

<p>Estimating the green view index (GVI) from street-level panoramas in hot-arid streetscapes is difficult because sparse vegetation, strong sunlight, and deep shadows reduce class separability. This study evaluated GVI reliability by comparing raw-image and polygon-guided pipelines for clarifying tree-crown boundaries. Using 14 georeferenced panoramas along a 3-km pedestrian corridor in Al-Madinah al-Munawwarah, Saudi Arabia, we benchmarked three models (DeepLab-like (ResNet50-ED), a SegNet-style model, and a Random Forest classifier) and assessed image-level error, agreement, bias, dispersion, input sensitivity, and pixel-wise delineation. Raw-image pipelines produced unstable and biased GVI estimates in high-contrast, low-greenness scenes. With polygon guidance, predictions became more repeatable but remained positively biased relative to the manual reference. DeepLab-like (ResNet50-ED) achieved the strongest image-level agreement (mean absolute error (MAE) = 6.08 percentage points (pp); root mean square error (RMSE) = 6.13 pp; Spearman’s <i>ρ</i> = 0.319), whereas Random Forest achieved high pixel-wise overlap (Intersection over Union (IoU) = 0.94) but higher image-level error (MAE = 8.93 pp; RMSE = 10.58 pp). These results show that pixel-wise overlap alone does not establish unbiased GVI measurement under a near-zero reference regime. We conclude that modest annotation combined with agreement-centered validation can improve the defensibility of street-level greenery indicators for hot-arid corridors. Remaining constraints include limited spatial and seasonal scope and annotation effort, motivating bias calibration, domain adaptation, and semi-supervised learning for scalable deployment.</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

An AI-based urban environmental monitoring framework for green view index assessment using street-level imagery in a hot-arid pedestrian corridor

  • 期刊论文

  • 发表日期：

    2026年7月10日

查看全文
  • Arwa AlSaedi

摘要

Estimating the green view index (GVI) from street-level panoramas in hot-arid streetscapes is difficult because sparse vegetation, strong sunlight, and deep shadows reduce class separability. This study evaluated GVI reliability by comparing raw-image and polygon-guided pipelines for clarifying tree-crown boundaries. Using 14 georeferenced panoramas along a 3-km pedestrian corridor in Al-Madinah al-Munawwarah, Saudi Arabia, we benchmarked three models (DeepLab-like (ResNet50-ED), a SegNet-style model, and a Random Forest classifier) and assessed image-level error, agreement, bias, dispersion, input sensitivity, and pixel-wise delineation. Raw-image pipelines produced unstable and biased GVI estimates in high-contrast, low-greenness scenes. With polygon guidance, predictions became more repeatable but remained positively biased relative to the manual reference. DeepLab-like (ResNet50-ED) achieved the strongest image-level agreement (mean absolute error (MAE) = 6.08 percentage points (pp); root mean square error (RMSE) = 6.13 pp; Spearman’s ρ = 0.319), whereas Random Forest achieved high pixel-wise overlap (Intersection over Union (IoU) = 0.94) but higher image-level error (MAE = 8.93 pp; RMSE = 10.58 pp). These results show that pixel-wise overlap alone does not establish unbiased GVI measurement under a near-zero reference regime. We conclude that modest annotation combined with agreement-centered validation can improve the defensibility of street-level greenery indicators for hot-arid corridors. Remaining constraints include limited spatial and seasonal scope and annotation effort, motivating bias calibration, domain adaptation, and semi-supervised learning for scalable deployment.