Scholarlink

<p>Accurate mapping of impervious surface area (ISA) is essential for urban planning and environmental monitoring. This study evaluates twelve spectral indices and two machine learning classifiers, namely Support Vector Machine (SVM) and Random Trees (RT), for ISA classification across four seasons in Taipei City using Sentinel-2A imagery. Classification accuracy, spectral separability, and ISA estimates were assessed with reference to Google Earth’s Dynamic World dataset. Among the indices, OSAVI, VrNIR-BI, and BLFEI performed best, with OSAVI and VrNIR-BI combined with SVM achieving the highest mean accuracies of 95.5% and 95.0% respectively across seasons. Seasonal variation in accuracy was minimal, and analysis of variance (ANOVA) results confirmed no statistically significant differences across seasons (<i>p</i> &gt; 0.05). ISA estimates were consistently lower than those from Dynamic World, which tended to overestimate built-up areas. SVM achieved slightly higher mean accuracies than RT, particularly when paired with well-performing indices. These findings underscore the value of integrating robust indices with SVM for accurate, seasonally consistent ISA classification and highlight the benefits of tailored index–classifier combinations in subtropical urban environments.</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

Performance of Spectral Indices and Machine Learning Algorithms in Seasonal Classification of Urban Impervious Surfaces from Sentinel-2 Imagery: A Case Study of Taipei

  • 期刊论文

  • 发表日期：

    2025年12月23日

查看全文
  • Alex Ssewanyana,
  • Min-Cheng Tu

摘要

Accurate mapping of impervious surface area (ISA) is essential for urban planning and environmental monitoring. This study evaluates twelve spectral indices and two machine learning classifiers, namely Support Vector Machine (SVM) and Random Trees (RT), for ISA classification across four seasons in Taipei City using Sentinel-2A imagery. Classification accuracy, spectral separability, and ISA estimates were assessed with reference to Google Earth’s Dynamic World dataset. Among the indices, OSAVI, VrNIR-BI, and BLFEI performed best, with OSAVI and VrNIR-BI combined with SVM achieving the highest mean accuracies of 95.5% and 95.0% respectively across seasons. Seasonal variation in accuracy was minimal, and analysis of variance (ANOVA) results confirmed no statistically significant differences across seasons (p > 0.05). ISA estimates were consistently lower than those from Dynamic World, which tended to overestimate built-up areas. SVM achieved slightly higher mean accuracies than RT, particularly when paired with well-performing indices. These findings underscore the value of integrating robust indices with SVM for accurate, seasonally consistent ISA classification and highlight the benefits of tailored index–classifier combinations in subtropical urban environments.