<![CDATA[Air pollution, particularly fine particulate matter (PM₂.₅), poses a critical threat to public health in rapidly urbanizing regions. Reliable early-warning systems are essential for mitigating exposure risks, yet challenges remain in cities with heterogeneous sensor coverage and event frequency. This study aimed to develop and evaluate a probabilistic, portable across cities early-warning framework for PM₂.₅ exceedances in Southeast Asia, focusing on Jakarta, Singapore, and Bangkok. Using a staged experimental design (Exp-A through Exp-E), we integrated regression-based back-casts with classification-based exceedance alerts, applied variant selection across thresholds and horizons, and validated for operational readiness through model freezing, documentation, and online simulation. Results showed that Jakarta achieved near-perfect exceedance prediction up to 6-hour horizons (F1 ≈ 0.99), Singapore maintained strong performance at short horizons (F1 ≈ 0.91 at 2–3 hours), while Bangkok yielded moderate but actionable signals at very short horizons (F1 ≈ 0.62 at 1 hour). Regression components provided stable situational awareness, and online smoothing reduced false alarms by approximately 15–20% without degrading performance. The framework demonstrated that calibrated exceedance probabilities can serve as an effective basis for city-level air quality alerts, with reliability strongly influenced by data density and event prevalence. This work contributes a reproducible, transparent, and computationally efficient approach that bridges machine learning innovation with practical environmental management. The findings emphasize the importance of horizon-specific calibration and adaptive strategies, offering both theoretical insights and practical value for policymakers in urban air quality governance.]]>
