Machine-Learning Proxy of Air-Quality Variability Associated with Intercontinental Bridge Traffic over the Bosphorus in Istanbul

Traffic-based air quality proxy models that perform well in compact urban settings have not been evaluated in megacity environments where complex topography, sea–land breeze circulation, and strong meteorological modulation are expected to dominate pollutant transport. We evaluate this transferability for the Bosphorus, an intercontinental corridor between Asia and Europe, using a 17,520-h panel for 2022–2023 that combines hourly traffic counts from six traffic sensors on the 15 Temmuz and Fatih Sultan Mehmet bridges with the closest air quality and meteorological observations. A traffic-proxy baseline (Prophet decomposition with a 9th-degree polynomial for hourly traffic flow) fails uniformly across all bridge × pollutant combinations, yielding negative coefficients of determination. A meteorology-aware machine-learning pipeline (LightGBM, XGBoost, and a bidirectional LSTM-Attention model) trained on traffic, meteorology, temporal indicators, precipitation lags, and short autoregressive terms recovers consistently positive predictive skill for NO2, NO, NOx, PM10, PM2.5, and the derived coarse-particle channel across both bridges. The mean of the twelve per-combination hold-out R2 values is ≈0.77 (range 0.58–0.86), yielding a 47–63% reduction in RMSE relative to the baseline; the three model families are practically equivalent in overall performance. SHAP ranks temporal autocorrelation, seasonal contrast, wind direction, and heavy vehicles (rather than total traffic) as the most influential predictors, whereas precipitation-delay characteristics carry diagnostic weight for particle channels. Distribution-free 95% prediction intervals are provided through split-conformal calibration. These results indicate that atmospheric persistence dominates predictive skill, with meteorology and traffic as secondary but systematic modulators.