Wildfire-Induced air pollution forecasting: Multi-Pollutant feature assessment and deep recurrent sequence learning for extreme smoke episodes

Environmental Modeling & Assessment · Springer Nature · 2026-11-09 · DOI: 10.1007/s11869-026-02155-6

Аннотация

Wildfire-induced biomass burning and tropical peatland smoke present catastrophic public health hazards and operational challenges for atmospheric forecasting in fire-prone tropical landscapes. Proactive, high-resolution air quality forecasting provides indispensable operational lead-time for health authorities to enact early warning advisories and mitigate acute respiratory and cardiovascular morbidity. However, standard machine learning frameworks frequently suffer from severe under-prediction plateaus during extreme fine particulate matter ($$\text {PM}_{2.5}$$) episodes due to non-linear chemical dynamics and out-of-distribution extrapolation constraints. In this study, we present an anomaly-aware deep learning framework tailored for short-term $$\text {PM}_{2.5}$$ forecasting during extreme haze episodes in West Kalimantan, Indonesia, a critical global epicenter for smoldering tropical peat combustion. Utilizing a multi-year telemetry dataset of 80,400 hourly observations (January 2021 to August 2025) across eight ground-based Air Quality Index ( Indeks Standar Pencemar Udara , ISPU) monitoring stations, an Isolation Forest preprocessing pipeline was implemented to systematically filter structural sensor noise while preserving legitimate high-amplitude pollution signals. Feature importance assessment via tree-based ensembles established Carbon Monoxide (CO) as the primary chemical tracer ($$\approx 69\%$$ predictive contribution) for smoldering peat smoke. Comparative benchmarking revealed that while Linear Regression and Random Forest failed to track elevated smoke accumulations due to rigid prediction plateaus, the Deep Long Short-Term Memory (LSTM) architecture effectively captured non-linear accumulation dynamics across severe haze levels ($$\text {PM}_{2.5} > 150$$ ISPU index), achieving superior forecasting accuracy ($$R^2 = 0.89$$, $$\text {MAE} = 8.12$$, $$\text {RMSE} = 11.93$$). These findings demonstrate that coupling ground-based multi-pollutant chemical signatures with deep recurrent sequence learning provides a highly reliable, computationally efficient decision support system for real-time hazard warnings and regional public health protection.

Читать у издателя →

Похожие статьи