Frontiers in Environmental Science (Nov 2023)

Evaluating air quality and criteria pollutants prediction disparities by data mining along a stretch of urban-rural agglomeration includes coal-mine belts and thermal power plants

  • Arti Choudhary,
  • Arti Choudhary,
  • Pradeep Kumar,
  • Chinmay Pradhan,
  • Saroj K. Sahu,
  • Sumit K. Chaudhary,
  • Pawan K. Joshi,
  • Pawan K. Joshi,
  • Deep N. Pandey,
  • Divya Prakash,
  • Divya Prakash,
  • Ashutosh Mohanty

DOI
https://doi.org/10.3389/fenvs.2023.1132159
Journal volume & issue
Vol. 11

Abstract

Read online

Air pollution has become a threat to human life around the world since researchers have demonstrated several effects of air pollution to the environment, climate, and society. The proposed research was organized in terms of National Air Quality Index (NAQI) and air pollutants prediction using data mining algorithms for particular timeframe dataset (01 January 2019, to 01 June 2021) in the industrial eastern coastal state of India. Over half of the study period, concentrations of PM2.5, PM10 and CO were several times higher than the NAQI standard limit. NAQI, in terms of consistency and frequency analysis, revealed that moderate level (ranges 101–200) has the maximum frequency of occurrence (26–158 days), and consistency was 36%–73% throughout the study period. The satisfactory level NAQI (ranges 51–100) frequency occurrence was 4–43 days with a consistency of 13%–67%. Poor to very poor level of air quality was found 13–50 days of the year, with a consistency of 9%–25%. Random Forest (RF), Support Vector Machine (SVM), Bagged Multivariate Adaptive Regression Splines (MARS) and Bayesian Regularized Neural Networks (BRNN) are the data mining algorithms, that showed higher efficiency for the prediction of PM2.5, PM10, NO2 and SO2 except for CO and O3 at Talcher and CO at Brajrajnagar. The Root Mean Square Error (RMSE) between observed and predicted values of PM2.5 (ranges 12.40–17.90) and correlation coefficient (r) (ranges 0.83–0.92) for training and testing data indicate about slightly better prediction of PM2.5 by RF, SVM, bagged MARS, and BRNN models at Talcher in comparison to PM2.5 RMSE (ranges 13.06–21.66) and r (ranges 0.64–0.91) at Brajrajnagar. However, PM10 (RMSE: 25.80–43.41; r: 0.57–0.90), NO2 (RMSE: 3.00–4.95; r: 0.42–0.88) and SO2 (RMSE: 2.78–5.46; r: 0.31–0.88) at Brajrajnagar are better than PM10 (RMSE: 35.40–55.33; r: 0.68–0.91), NO2 (RMSE: 4.99–9.11; r: 0.48–0.92), and SO2 (RMSE: 4.91–9.47; r: 0.20–0.93) between observed and predicted values of training and testing data at Talcher using RF, SVM, bagged MARS and BRNN models, respectively. Taylor plots demonstrated that these algorithms showed promising accuracy for predicting air quality. The findings will help scientific community and policymakers to understand the distribution of air pollutants to strategize reduction in air pollution and enhance air quality in the study region.

Keywords