Scholarlink

<p>Focusing on the core area within Xi’an’s Second Ring Road, we integrate built-environment, street-view, and human-activity data and propose an explainable pipeline (TabPFN–SHAP → K-means, k = 4) to uncover nonlinear effects and interactions on land surface temperature (LST) and to delineate differentiated governance units. In model benchmarking, TabPFN achieves the best generalization on the test set (R² = 0.650 ± 0.038, MAE = 0.871 ± 0.041, MSE = 1.482 ± 0.153). Tree-based models fit the training data extremely well (R² ≥ 0.99) but generalize poorly (test R² = 0.326–0.415; ΔR² ≥ 0.25), indicating overfitting. SHAP global importance shows built-environment factors dominate: building density (BD = 31.8%) strongly warms LST; building height (BH = 18.6%) cools; road density (RD = 11.6%) has an inverted-U sensitivity (≈ 800–1300). Population density (PD = 13.8%) turns positively related at high levels. In street views, the green view index (GVI = 11.5%) is strongly negative to LST with a threshold ≥ 0.15 and a cooling band of 0.20–0.35; the sky view index (SVI = 5.2%) is weakly positive; higher IEI and BCR intensify heating. K-means yields four thermal clusters validated by spatial autocorrelation (Moran’s I = 0.303, z = 21.02, <i>p</i> &lt; 0.001), with net SHAP effects of − 1.10, + 0.06, + 0.20, and + 2.83, corresponding to cool-island, mildly warming, commuter-heating, and old-city/commercial-core intensifying types. We translate these into actionable rules: along-street GVI ≥ 0.15 (target 0.20–0.35); BH ≥ 60–100 in high-mass districts to ensure continuous shading and ventilation; “open connections + continuous canopies” where IEI &gt; 0.5; and an infill control line at BCR ≈ 0.20.</p> Graphical abstract <p></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

Study on the influence mechanism of urban environmental factors on thermal environment based on machine learning method

  • 期刊论文

  • 发表日期：

    2026年4月14日

查看全文
  • Yonghao Li,
  • Yuyi Chen

摘要

Focusing on the core area within Xi’an’s Second Ring Road, we integrate built-environment, street-view, and human-activity data and propose an explainable pipeline (TabPFN–SHAP → K-means, k = 4) to uncover nonlinear effects and interactions on land surface temperature (LST) and to delineate differentiated governance units. In model benchmarking, TabPFN achieves the best generalization on the test set (R² = 0.650 ± 0.038, MAE = 0.871 ± 0.041, MSE = 1.482 ± 0.153). Tree-based models fit the training data extremely well (R² ≥ 0.99) but generalize poorly (test R² = 0.326–0.415; ΔR² ≥ 0.25), indicating overfitting. SHAP global importance shows built-environment factors dominate: building density (BD = 31.8%) strongly warms LST; building height (BH = 18.6%) cools; road density (RD = 11.6%) has an inverted-U sensitivity (≈ 800–1300). Population density (PD = 13.8%) turns positively related at high levels. In street views, the green view index (GVI = 11.5%) is strongly negative to LST with a threshold ≥ 0.15 and a cooling band of 0.20–0.35; the sky view index (SVI = 5.2%) is weakly positive; higher IEI and BCR intensify heating. K-means yields four thermal clusters validated by spatial autocorrelation (Moran’s I = 0.303, z = 21.02, p < 0.001), with net SHAP effects of − 1.10, + 0.06, + 0.20, and + 2.83, corresponding to cool-island, mildly warming, commuter-heating, and old-city/commercial-core intensifying types. We translate these into actionable rules: along-street GVI ≥ 0.15 (target 0.20–0.35); BH ≥ 60–100 in high-mass districts to ensure continuous shading and ventilation; “open connections + continuous canopies” where IEI > 0.5; and an infill control line at BCR ≈ 0.20.

Graphical abstract