Title : Explainable machine learning and integrated statistical assessment of urban air quality dynamics using multi-pollutant monitoring data of Thankot Corridor, Kathmandu Valley
Abstract:
Urban air pollution continues to be an environmental and health issue for fast-urbanizing areas in South Asia, and the Kathmandu Valley in Nepal is particularly at risk due to the valley’s bowllike geographic nature, which does not allow for effective dispersal of contaminants in the atmosphere. While previous research has noted higher levels of contaminants in the valley, the literature lacks a complete and well-developed study that uses robust statistical analysis combined with explainable machine learning, specifically in highly traveled areas like Thankot, which serves as an important western entrance to Kathmandu as well as to the whole country of Nepal. In order to fill the gap, the current research provides an analysis of air quality dynamics through the use of both statistical analysis and explainable machine learning methods applied to a multi-pollutant data set including 2,420 measurements of ten different pollutants and meteorological parameters: PM2.5, PM10, PM1.0, CO2, formaldehyde, TVOC, temperature, wind speed, relative humidity, and ambient sound level. Descriptive statistics, Pearson correlation, Shapiro-Wilk tests for normality, non-parametric Mann-Kendall trend test with Sen’s slope, and principal component analysis were used to analyze pollutant variability, temporal trends, and multivariate dominant patterns. The Random Forest Regression Model compared to Multiple Linear Regression Models and validated using 80/20 train-test splits and fivefold cross-validation was developed to estimate the PM2.5 concentrations using co-measured pollutants and meteorological factors. The SHAP values combined with ICE plots were utilized to explain the behavior of the models and evaluate the contribution of the individual predictors. It was found that the mean PM2.5 concentrations were equal to 41.8 ± 17.1 μg/m³ which is nearly eight times higher than the WHO guideline and the highest concentration was detected at the station located at the busiest intersection. All pollutants demonstrated statistically significant decreasing trends (p < 0.001) throughout the study period due to the change of seasons from dry to monsoon and associated with scavenging of pollutants due to increased humidity. Through principal component analysis, the ten-variable system was transformed into two meaningful axes – namely, pollutant load and meteorological factors – which accounted for 68.7% of the total variance. In turn, the Random Forest model provided highly predictive results with an R² value of 0.878, RMSE of 5.97 μg/m³ and MAE of 4.81 μg/m³, slightly superior to the linear model, while the SHAP analysis revealed TVOC and relative humidity to be the main factors driving the PM2.5 concentrations with formaldehyde and ambient noise as secondary traffic-related variables. Overall, these results illustrate how the integration of scientific statistics and explainable machine learning can help in characterizing the behavior of urban air pollution, providing a solid scientific base for making decisions on air-quality management, public health protection, and sustainable environmental policies in the Kathmandu Valley.
