A support vector machine classifier of PM₂.₅-based air quality index categories from co-pollutants and meteorological variables in Ahvaz, Iran

Ahvaz, in southwestern Iran, experiences some of the most severe particulate air pollution in the world, driven by Middle Eastern Dust events, industrial emissions, and traffic. Continuous monitoring of fine particulate matter (PM₂.₅) is essential for public-health protection, yet PM₂.₅ monitors in the region frequently suffer outages and data gaps. This study evaluates whether the PM₂.₅-based Air Quality Index (AQI) category can be estimated from routinely measured co-pollutants (O₃, NO, NO₂, SO₂, CO, and PM₁₀) and meteorological variables (air temperature, dew point, relative humidity, wind speed, and wind direction), providing a surrogate risk label when direct PM₂.₅ measurements are unavailable. Hourly observations from the Naderi monitoring station in Ahvaz (October 2014 – July 2017) were aggregated into 3-hour intervals; each interval with at least two of three observed hourly PM₂.₅ values was labelled with one of six AQI categories using the U.S. EPA PM₂.₅ breakpoints in force during the study period, yielding 7,509 samples. A class-weighted radial-basis-function Support Vector Machine (SVM) was tuned by rolling-origin cross-validation within the training period (October 2014 – January 2017) and evaluated on a strictly held-out chronological test period (January – July 2017), alongside multinomial logistic regression, linear SVM, random forest, and gradient boosting baselines trained under identical protocols. On the unseen test period, the SVM achieved an exact-category accuracy of 0.633, macro-averaged F1 of 0.38, Cohen’s κ of 0.47, and quadratic-weighted κ of 0.61; 94.5% of predictions fell within one category of the truth, and misclassifications concentrated between adjacent categories. The random forest baseline performed best overall (accuracy 0.674; quadratic-weighted κ of 0.70; 97.7% within one category). Permutation importance and ablation experiments identified PM₁₀ as the dominant predictor: excluding it reduced SVM accuracy from 0.633 to 0.431, while PM₁₀ alone reached 0.523. Performance for the rare extreme categories (Very Unhealthy and Hazardous) remained limited and is reported transparently. At this station and over this period, co-pollutant and meteorological measurements—chiefly PM₁₀—can provide an operationally useful, near-real-time surrogate for the PM₂.₅ AQI category when the PM₂.₅ channel fails, while underscoring that such a surrogate cannot replace properly maintained PM₂.₅ instrumentation, particularly for extreme pollution episodes.