Point process modeling of wildfire hazard in Los Angeles County, California

TL;DR

This study evaluates advanced point process models incorporating weather and spatial data to improve wildfire hazard prediction in Los Angeles County, outperforming the traditional Burning Index.

Contribution

It introduces complex point process models with spatial and weather covariates that significantly enhance wildfire forecasting accuracy over existing BI-based methods.

Findings

Multiplicative models with weather variables improve fit.

Spatial bandwidth parameters enhance model performance.

Models outperform BI in predicting wildfires.

Abstract

The Burning Index (BI) produced daily by the United States government's National Fire Danger Rating System is commonly used in forecasting the hazard of wildfire activity in the United States. However, recent evaluations have shown the BI to be less effective at predicting wildfires in Los Angeles County, compared to simple point process models incorporating similar meteorological information. Here, we explore the forecasting power of a suite of more complex point process models that use seasonal wildfire trends, daily and lagged weather variables, and historical spatial burn patterns as covariates, and that interpolate the records from different weather stations. Results are compared with models using only the BI. The performance of each model is compared by Akaike Information Criterion (AIC), as well as by the power in predicting wildfires in the historical data set and residual analysis. We find that multiplicative models that directly use weather variables offer substantial improvement in fit compared to models using only the BI, and, in particular, models where a distinct spatial bandwidth parameter is estimated for each weather station appear to offer substantially improved fit.