Scholarlink

<p>Mapping urban features is essential for better planning, environmental management, and fair use of resources, particularly when dealing with Blue–Green Infrastructure (BGI). While many studies use machine learning to improve classification, we still do not fully understand how different spectral bands behave in varied cities, especially in complex and rapidly changing urban environments. Therefore, in this research, we present a novel framework for BGI classification and accuracy enhancement using Game theory SHapley Additive exPlanations (SHAP) values integrating with machine learning techniques and Sentinel-2 satellite image for scalability of the methodology. In order to examine blue and green infrastructures at city-level, this study examines game theory SHAP values integrating with random forest (RF) model used in the study. Using a large dataset of 57,221 data points, the performance of the model is methodically assessed, with a focus on improving interpretability by utilizing SHAP values. Beyond conventional accuracy assessments, this research investigates the subtle spectral characteristics influencing the recognition of infrastructure across distinct classes, including Deep Green, Green, and Blue. Notably, spectral bands such as B8A, B7, B6, and B8 demonstrate remarkable precision in categorizing the Deep Green class, while the importance of bands like B8A, B5, B4, and B3 is showed importance distinguishing green infrastructures. Additionally, the combination of bands B5, B4, B3, and B12B5 demonstrate as an important discriminator for accurately identifying blue infrastructures. We enhanced infrastructure classification through granular analysis utilizing Shapley values, offering insights into the importance of specific spectral bands for classification outcomes. These findings convey significant implications for urban planners, environmentalists, and policymakers, offering valuable insights into optimizing urban feature classification accuracy and refining models for precise infrastructure mapping.</p> Graphical Abstract <p>The graphical abstract illustrates the workflow and methodology employed for classifying blue and green infrastructure (BGI) in the Bangkok Metropolitan Region using Sentinel-2 satellite imagery and advanced machine learning techniques. The process begins with data input, where a large multispectral satellite dataset consisting of 57,221 data points is collected to capture spatial and spectral variations across the urban landscape. These inputs are then processed using a Random Forest categorization model, which leverages machine learning-based clustering algorithms to classify land cover into Deep Green, Green, and Blue categories. Model validation is performed using Out-of-Bag (OOB) accuracy evaluation, ensuring reliable assessment of classification performance without requiring an independent validation set. The adopted methods section highlights three complementary approaches to enhance interpretability and accuracy: SHAP (SHapley Additive exPlanations) values, band cross-collinearity analysis, and Principal Component Analysis (PCA). SHAP provides a transparent view of the contribution of individual spectral bands to classification decisions, while cross-collinearity analysis identifies linear relationships among bands, and PCA reduces data dimensionality to capture maximum variance. These techniques collectively enable the. identification of the best-performing spectral bands, which are crucial for distinguishing between different infrastructure types. Finally, the results and outputs section presents the classification maps and highlights the optimal bands for BGI classification. The map outputs reveal the spatial distribution of Deep Green, Green, and Blue infrastructure in Bangkok, providing actionable insights for urban planners. By integrating accurate classification with interpretable feature analysis, the framework supports resilience through sustainable infrastructure planning, informing policies to enhance green-blue connectivity, urban ecosystem services, and climate adaptation strategies in densely built environments. This graphical abstract encapsulates the synergy of satellite data, machine learning, and explainable AI for sustainable urban management.</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

Improving Urban Landscape Mapping with SHAP for Optimized Blue-Green Infrastructure Classification in Tropical Cities

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Md. Mehedi Hasan,
  • Malay Pramanik,
  • Swapan Talukdar,
  • Iftekharul Alam,
  • Atul Diwakar,
  • Bijay Halder,
  • Kanak N. Moharir,
  • Chaitanya Baliram Pande,
  • Mohamed Zhran

摘要

Mapping urban features is essential for better planning, environmental management, and fair use of resources, particularly when dealing with Blue–Green Infrastructure (BGI). While many studies use machine learning to improve classification, we still do not fully understand how different spectral bands behave in varied cities, especially in complex and rapidly changing urban environments. Therefore, in this research, we present a novel framework for BGI classification and accuracy enhancement using Game theory SHapley Additive exPlanations (SHAP) values integrating with machine learning techniques and Sentinel-2 satellite image for scalability of the methodology. In order to examine blue and green infrastructures at city-level, this study examines game theory SHAP values integrating with random forest (RF) model used in the study. Using a large dataset of 57,221 data points, the performance of the model is methodically assessed, with a focus on improving interpretability by utilizing SHAP values. Beyond conventional accuracy assessments, this research investigates the subtle spectral characteristics influencing the recognition of infrastructure across distinct classes, including Deep Green, Green, and Blue. Notably, spectral bands such as B8A, B7, B6, and B8 demonstrate remarkable precision in categorizing the Deep Green class, while the importance of bands like B8A, B5, B4, and B3 is showed importance distinguishing green infrastructures. Additionally, the combination of bands B5, B4, B3, and B12B5 demonstrate as an important discriminator for accurately identifying blue infrastructures. We enhanced infrastructure classification through granular analysis utilizing Shapley values, offering insights into the importance of specific spectral bands for classification outcomes. These findings convey significant implications for urban planners, environmentalists, and policymakers, offering valuable insights into optimizing urban feature classification accuracy and refining models for precise infrastructure mapping.

Graphical Abstract

The graphical abstract illustrates the workflow and methodology employed for classifying blue and green infrastructure (BGI) in the Bangkok Metropolitan Region using Sentinel-2 satellite imagery and advanced machine learning techniques. The process begins with data input, where a large multispectral satellite dataset consisting of 57,221 data points is collected to capture spatial and spectral variations across the urban landscape. These inputs are then processed using a Random Forest categorization model, which leverages machine learning-based clustering algorithms to classify land cover into Deep Green, Green, and Blue categories. Model validation is performed using Out-of-Bag (OOB) accuracy evaluation, ensuring reliable assessment of classification performance without requiring an independent validation set. The adopted methods section highlights three complementary approaches to enhance interpretability and accuracy: SHAP (SHapley Additive exPlanations) values, band cross-collinearity analysis, and Principal Component Analysis (PCA). SHAP provides a transparent view of the contribution of individual spectral bands to classification decisions, while cross-collinearity analysis identifies linear relationships among bands, and PCA reduces data dimensionality to capture maximum variance. These techniques collectively enable the. identification of the best-performing spectral bands, which are crucial for distinguishing between different infrastructure types. Finally, the results and outputs section presents the classification maps and highlights the optimal bands for BGI classification. The map outputs reveal the spatial distribution of Deep Green, Green, and Blue infrastructure in Bangkok, providing actionable insights for urban planners. By integrating accurate classification with interpretable feature analysis, the framework supports resilience through sustainable infrastructure planning, informing policies to enhance green-blue connectivity, urban ecosystem services, and climate adaptation strategies in densely built environments. This graphical abstract encapsulates the synergy of satellite data, machine learning, and explainable AI for sustainable urban management.