Journal Article

Advanced Machine Learning Techniques for Precise hourly Air Quality Index (AQI) Prediction in Azamgarh, India

Ansari, Asif, Quaff, Abdur Rahman

2025

Overview

This paper forecasts the hourly AQI in Azamgarh, Uttar Pradesh, India, using machine learning (ML) models. This work involved the real-time measurement of hourly particulate matter (PM 2.5 , PM 10 ), gaseous concentrations (NO 2 , SO 2 ), and meteorological parameters (temperature, relative humidity, wind direction, wind speed, and uv radiation) in the ambient air. We have calculated the pollutant subindex and hourly AQI for Azamgarh from July 2022 to June 2023, encompassing a total of 8760 data points. The study observed annual average concentrations of NO 2 (33.91 ± 30.38 µg/m 3 ), SO 2 (43.96 ± 20.12 µg/m 3 ), PM 2.5 (52.25 ± 55.26 µg/m 3 ), and PM 10 (77.58 ± 80.61 µg/m 3 ). Meteorological conditions included a mean temperature of 27.20 ± 7.65 °C, relative humidity of 63.93 ± 18.74%, wind speed of 1.00 ± 0.96 m/s, and a uv index of 0.42 ± 0.34. The computed annual mean hourly AQI was 123 ± 91.96, which falls into the moderate-level pollution category. The AQI values in winter were higher than in summer. ANOVA analysis shows statistically significant monthly changes in pollutants and meteorological parameters: PM 2.5 (F = 842.7), PM 10 (F = 773.5), NO 2 (F = 583.6), SO 2 (F = 3356), temperature (F = 1045), humidity (F = 572.5), wind speed (F = 36.9), wind direction (F = 54.27), uv index (F = 24.48), and hourly AQI (F = 738.4), all with p < 2e-16. We predicted the hourly AQI using eight machine learning models: adaboost, catboost, gradient boosting, knn, linear regression, random forest, svm, and xgboost. With the fastest computational time of 1.61 s out of all the models examined, xgboost stood out as the best model because of its outstanding R 2 and RMSE performances. We employed a tenfold cross-validation to evaluate the models' performance. The results of the paired t-test indicate that catboost, gradboost, linear regression, and xgboost statistically align closely with the observed data. We can also observe the models' comparative performance using a taylor diagram. An investigation of sensitivity revealed that PM 2.5 , PM 10 , NO 2 , and SO 2 are crucial for the accuracy of the xgboost model in predicting AQI. Excluding these pollutants greatly reduces performance. In conclusion, machine learning can predict AQI, and xgboost is the best model for this purpose. Graphical Abstract Highlights Hourly gaseous, particulate and metrological components were measured (8760 data points). To forecast hourly AQI, we employed eight machine-learning algorithms. Annual mean AQI is 123 (moderate); higher in winter. XGboost stood out as the best models because of their outstanding R-squared, RMSE and least computational time (1.61 s). Cross-validation (tenfold) and Taylor diagram was employed to evaluate the models' performance. Sensitivity analysis reveals PM 2.5 , PM 10 , NO 2 , and SO 2 are crucial for XGboost's AQI predictions.

• Publisher: Springer International Publishing,Springer Nature B.V
• Subject: Air pollution / Air quality / Computer applications / Computing time / Data points / Earth and Environmental Science / Environment / Environmental Engineering/Biotechnology / Environmental Management / Geoecology/Natural Processes / Landscape/Regional and Urban Planning / Learning algorithms / Machine learning / Meteorological parameters / Mortality / Natural Hazards / Neural networks / Nitrogen dioxide / Outdoor air quality / Particulate emissions / Particulate matter / Performance evaluation / Pollutants / Pollution control / Real time / Relative humidity / Research Paper / Sensitivity analysis / Statistical analysis / Sulfur dioxide / Ultraviolet radiation / Variance analysis / Wind direction / Wind speed / Winter