Abstract
Tropical peatland and forest fires are critical contributors to regional haze and public health crises, yet ground-based monitoring in these regions remains sparse. This study develops a robust framework for estimating surface PM₂.₅ concentrations during extreme fire events (2021–2025) by integrating multi-sensor satellite observations with in-situ data through a Hist Gradient Boosting Regressor (HGBR) approach. To enhance predictive accuracy, we implemented advanced feature engineering, including 1–3 days of exogenous lags, rolling statistics (3 and 7-day windows), and aerosol–meteorological interaction variables. Our analysis of multiple pollutants (PM₂.₅, NO₂, SO₂, CO, HC, and O₃) reveals that during active fire periods, the Air Quality Index (AQI) frequently escalated to "Unhealthy" and "Hazardous" levels. The proposed HGBR model demonstrated high fidelity in representing spatiotemporal variability, achieving a coefficient of determination (R² = 0.72), with an RMSE of 14.26 μg/m³ and MAE of 8.49 μg/m³ (n = 339). These results validate the efficacy of machine learning-driven satellite monitoring in bypassing the limitations of fragmented ground station networks. This framework offers a scalable solution for operational air quality forecasting and early warning systems in fire-prone equatorial regions.