Scholarlink

<p>Particulate matter (PM), particularly those with an aerodynamic diameter of less than 2.5µm (PM<sub>2.5</sub>), is linked to various health issues, including respiratory and cardiovascular diseases. With its dense population, extensive diesel-dominated vehicle fleet, expanding commercial sectors, and numerous industrial facilities, Kolkata faces severe air quality challenges. The city frequently experiences ‘very unhealthy’ AQI levels, primarily driven by elevated PM<sub>2.5</sub> concentrations during winter. Despite the availability of ground-based PM measurements, their spatial and temporal limitations hinder comprehensive air quality assessment, especially in a rapidly urbanizing city like Kolkata. This study estimates ground-level PM<sub>2.5</sub> and PM<sub>10</sub> concentrations from 2015 to 2022 using Aerosol Optical Depth (AOD) derived from MODIS MOD04_L2 product with a spatial resolution of 10&#xa0;km. This study also examines the influence of meteorological variables (temperature, relative humidity, precipitation, wind speed, and cloud cover) on the AOD-PM relationship. Two machine learning models were developed: an Artificial Neural Network (ANN) with three hidden layers optimized through backpropagation and a Random Forest Regression (RF) model utilizing ensemble learning. Model performance was validated using ground-based PM and independent test data from October 2023. Quantitative results show that both models performed well in predicting PM concentrations. For PM<sub>2.5</sub>, the RF model achieved an R<sup>2</sup> of 0.961 for training and 0.952 for testing, with a lower RMSE of 12.4&#xa0;µg/m³, while the ANN model achieved an R<sup>2</sup> of 0.97 for training and 0.95 for testing, with a higher RMSE of 14.8&#xa0;µg/m³. For PM<sub>10</sub>, the RF model produced an R<sup>2</sup> of 0.965 for training and 0.932 for testing, with an RMSE of 18.2&#xa0;µg/m³, while the ANN model showed an R² of 0.987 for training and 0.947 for testing, with an RMSE of 21.5&#xa0;µg/m³. Although both models demonstrated strong predictive capabilities, the RF model was selected due to its lower RMSE values and greater consistency across training and testing phases. Among meteorological factors, temperature emerged as the most significant predictor, followed by wind speed, in influencing PM concentrations. The findings demonstrate the successful integration of machine learning and geospatial technology for PM estimation, overcoming the limitations of ground-based measurements. While the study highlights location-specific challenges in Kolkata, the methods employed provide a transferable framework for improving air quality assessments in other urban regions globally.</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

Simulation of Ground-Level Particulate Matter Concentration with AOD and Related Meteorological Parameters: A Case Study on Kolkata by Random Forest and ANN

  • 期刊论文

  • 发表日期：

    2025年2月3日

查看全文
  • Eeshan Basu,
  • Chalantika Laha Salui,
  • Indranil Mukherjee,
  • Debasis Sau

摘要

Particulate matter (PM), particularly those with an aerodynamic diameter of less than 2.5µm (PM2.5), is linked to various health issues, including respiratory and cardiovascular diseases. With its dense population, extensive diesel-dominated vehicle fleet, expanding commercial sectors, and numerous industrial facilities, Kolkata faces severe air quality challenges. The city frequently experiences ‘very unhealthy’ AQI levels, primarily driven by elevated PM2.5 concentrations during winter. Despite the availability of ground-based PM measurements, their spatial and temporal limitations hinder comprehensive air quality assessment, especially in a rapidly urbanizing city like Kolkata. This study estimates ground-level PM2.5 and PM10 concentrations from 2015 to 2022 using Aerosol Optical Depth (AOD) derived from MODIS MOD04_L2 product with a spatial resolution of 10 km. This study also examines the influence of meteorological variables (temperature, relative humidity, precipitation, wind speed, and cloud cover) on the AOD-PM relationship. Two machine learning models were developed: an Artificial Neural Network (ANN) with three hidden layers optimized through backpropagation and a Random Forest Regression (RF) model utilizing ensemble learning. Model performance was validated using ground-based PM and independent test data from October 2023. Quantitative results show that both models performed well in predicting PM concentrations. For PM2.5, the RF model achieved an R2 of 0.961 for training and 0.952 for testing, with a lower RMSE of 12.4 µg/m³, while the ANN model achieved an R2 of 0.97 for training and 0.95 for testing, with a higher RMSE of 14.8 µg/m³. For PM10, the RF model produced an R2 of 0.965 for training and 0.932 for testing, with an RMSE of 18.2 µg/m³, while the ANN model showed an R² of 0.987 for training and 0.947 for testing, with an RMSE of 21.5 µg/m³. Although both models demonstrated strong predictive capabilities, the RF model was selected due to its lower RMSE values and greater consistency across training and testing phases. Among meteorological factors, temperature emerged as the most significant predictor, followed by wind speed, in influencing PM concentrations. The findings demonstrate the successful integration of machine learning and geospatial technology for PM estimation, overcoming the limitations of ground-based measurements. While the study highlights location-specific challenges in Kolkata, the methods employed provide a transferable framework for improving air quality assessments in other urban regions globally.