Predicting hourly PM_2.5 concentrations in wildfire-prone areas using a SpatioTemporal Transformer model

Area of Science:

• Environmental Science
• Data Science
• Atmospheric Science

Background:

• Wildfires are increasing globally, causing widespread air pollution from smoke.
• Accurate forecasting of particulate matter (PM2.5) is crucial for public health, especially in wildfire-prone regions.
• Existing air quality models struggle to predict localized PM2.5 spikes during wildfires.

Purpose of the Study:

• To develop an advanced deep learning model for improved spatiotemporal prediction of PM2.5 concentrations.
• To enhance the accuracy and interpretability of air pollution forecasts related to wildfire events.
• To provide timely and reliable air quality information to citizens affected by wildfire smoke.

Main Methods:

• Proposed a novel multi-head attention-based deep learning architecture, the SpatioTemporal (ST)-Transformer.
• Utilized a sparse attention mechanism to focus on relevant spatial, temporal, and variable-wise contextual information.
• Incorporated key factors like wildfire perimeters, intensity, meteorology, traffic, and historical PM2.5 data for predictions.

Main Results:

• The ST-Transformer demonstrated superior performance in predicting PM2.5 concentrations compared to existing time series methods.
• The model effectively captured abrupt changes and spikes in PM2.5 levels during wildfire events.
• Learned attention matrices provided interpretability, distinguishing between wildfire and non-wildfire scenarios.

Conclusions:

• The ST-Transformer offers accurate and interpretable spatiotemporal predictions for PM2.5, crucial for wildfire smoke impact assessment.
• This model can significantly aid in monitoring and predicting air quality during wildfire events.
• The architecture's adaptability suggests potential applications in other complex spatiotemporal prediction challenges.