Scholarlink

<p>Comprehending the various classifications of aerosols is essential for assessing their implications on the planet’s energy equilibrium and formulating robust climate policies. This investigation examines the efficacy of unsupervised machine learning methodologies to categorize aerosol types while exhibiting minimal reliance on annotated datasets. We present an innovative framework that integrates feature permutation alongside data balancing methodologies to augment the effectiveness of unsupervised algorithms, which include Agglomerative clustering, K-means clustering, Gaussian Mixture Model (GMM), and Balanced Iterative Reducing and Clustering using Hierarchies (BIRCH) clustering. This study employs aerosol characteristics derived using the Aerosol Robotic Network (AERONET), gathered over the year from 2001 to 2022 at three sites in North India: Jaipur, Gandhi College, and Kanpur, selected to capture diverse regional settings, including urban and Indo-Gangetic Plain environments, and owing to the availability of long-term, high-quality datasets. The research accurately identified seven to eight discrete aerosol types: Polluted Dust-Mixed Absorbing (MA), Mixed Non-Absorbing (MNA), Black Carbon Low (BCL), Black Carbon Medium (BCM), Black Carbon High (BCH), Dust-Coarse Absorbing (CA), and Marine-Coarse Non-Absorbing (CNA). We examined the cluster labelling capabilities of the algorithm using manually classified labels, making use of the thresholding method. The results revealed that the Jaipur site attained the highest accuracy at 94.68%, followed by the Gandhi College site with an accuracy of 92.00%, and the Kanpur site, which achieved an accuracy of 89.66%. The study additionally examined the seasonal variations of aerosol classifications within the three designated study areas. This study addresses a significant issue in aerosol classification by employing an unsupervised learning strategy that significantly lessens the need for labelled data while preserving excellent classification accuracy.</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

A Novel Framework Using Unsupervised Machine Learning for Aerosol Classification: A Case Study of North Indian AERONET Sites

  • 期刊论文

  • 发表日期：

    2026年6月22日

查看全文
  • Varsha Hole,
  • Manu Mehta,
  • Subhash Shinde,
  • Chaitya Dobariya,
  • Aayush Patel,
  • Arshad Mapari

摘要

Comprehending the various classifications of aerosols is essential for assessing their implications on the planet’s energy equilibrium and formulating robust climate policies. This investigation examines the efficacy of unsupervised machine learning methodologies to categorize aerosol types while exhibiting minimal reliance on annotated datasets. We present an innovative framework that integrates feature permutation alongside data balancing methodologies to augment the effectiveness of unsupervised algorithms, which include Agglomerative clustering, K-means clustering, Gaussian Mixture Model (GMM), and Balanced Iterative Reducing and Clustering using Hierarchies (BIRCH) clustering. This study employs aerosol characteristics derived using the Aerosol Robotic Network (AERONET), gathered over the year from 2001 to 2022 at three sites in North India: Jaipur, Gandhi College, and Kanpur, selected to capture diverse regional settings, including urban and Indo-Gangetic Plain environments, and owing to the availability of long-term, high-quality datasets. The research accurately identified seven to eight discrete aerosol types: Polluted Dust-Mixed Absorbing (MA), Mixed Non-Absorbing (MNA), Black Carbon Low (BCL), Black Carbon Medium (BCM), Black Carbon High (BCH), Dust-Coarse Absorbing (CA), and Marine-Coarse Non-Absorbing (CNA). We examined the cluster labelling capabilities of the algorithm using manually classified labels, making use of the thresholding method. The results revealed that the Jaipur site attained the highest accuracy at 94.68%, followed by the Gandhi College site with an accuracy of 92.00%, and the Kanpur site, which achieved an accuracy of 89.66%. The study additionally examined the seasonal variations of aerosol classifications within the three designated study areas. This study addresses a significant issue in aerosol classification by employing an unsupervised learning strategy that significantly lessens the need for labelled data while preserving excellent classification accuracy.