Do Satellite Active-Fire Detections Improve Short-Term PM2.5 Forecasts? A Chronological Ablation Study in Sammamish, Washington

Satellite wildfire data are often assumed to improve short-term PM2.5 forecasts, but their value relative to persistence remains uncertain. This study tests that assumption in Sammamish, Washington, using 6,882 sensor-days from calibrated PurpleAir measurements, Open-Meteo weather, and NASA FIRMS fire detections collected during 2021–2026. Three nested feature sets: PM2.5 history (M1), weather (M2), and wildfire activity (M3) were evaluated with linear regression, Random Forest, and XGBoost at one- and two-day horizons using chronological splits. Wildfire features did not consistently improve accuracy: M3 increased overall MAE in five of six model–horizon comparisons, while the best t+1 model improved on persistence by only 2.7%. On elevated-PM2.5 days, M3 reduced Random Forest’s t+2 MAE by 23.2%, from 8.361 to 6.418 µg/m³, but remained worse than persistence at 5.850 µg/m³. On wildfire-active days, M3 increased error for every model, by up to 33.8%. Models also underpredicted elevated episodes while overpredicting on wildfire-active days, despite assigning 38–82% of feature importance to fire variables. These findings show that fire detections describe regional source activity but not receptor-level smoke transport, motivating transport-aware features and mandatory persistence comparisons. The codes are available here - https://github.com/APSo0d/AQ-Research