Despite California being a major fire hotspot in the Americas, there is no extensive scientific analysis of operational fire spread models allowing analysis of their performance and drivers leading to model inaccuracies. Recent advances in technology have allowed monitoring the fire progression of most wildfires every 15 min in the United States through the National Fireguard Detections platform. This data, when available for use on a fire, provides unprecedented capabilities to analyze factors influencing fire behavior and compare the observed and predicted wildfire rate-of-spread (ROS) modeling in fires distributed across different and complex landscapes.
Building on other studies that analyzed these modeling techniques, Technosylva joined with CALFIRE and led a 2023 peer-reviewed study, published in the International Journal of Wildland Fire that assesses the performance of fire spread models used in California by comparing observed fire growth data with simulated data. The analysis reviewed operational settings under different environmental conditions using 1853 California wildfires from 2019 to 2021 to determine what conditions the current models may over, or underestimate ROS and subsequently, the burned area and associated fire impacts on buildings and other assets.
”It was a great opportunity to analyze these fires because it's the first time we have had such a data set with its huge number of files and additionally, temporal resolution of that data in polygons every 15 minutes. So, it is unprecedented to have both this amount of fire monitoring data and a fire behavior simulator platform with high-quality inputs, including the fuel types, the weather conditions, canopy characteristics, and other pieces.
The analysis allowed us the opportunity to compare the best fire modeling possible with the best fire monitoring possible. The main conclusion from the analysis was that these models can be used in wildfire operational environments.
Adrián Cardil, Ph. DLead Author & Senior Fire Researcher
Insight from the Research
Wildfire spread models play a crucial role in predicting how fires propagate, but their accuracy is influenced by various factors, including fuel availability, topography, and weather. Among these models, Rothermel’s semi-empirical model has been widely used for its simplicity and computational efficiency. However, the inherent limitations and assumptions of these models, along with input data quality, can impact their reliability.
This study, conducted in California, aimed to assess the predictive accuracy of wildfire spread models under different environmental conditions. It utilized high-resolution data from the National Fireguard Detections product to compare observed and predicted Rates of Spread (ROS) for 1853 wildfires occurring from 2019 to 2021. The analysis sought to identify conditions under which the models overestimate or underestimate ROS, ultimately affecting the burned area and fire impacts on buildings and assets.
Cite: Adrián Cardil
Key observations and findings from the Research
- Fire Progression Data: The study used the National Fireguard Detections product data, offering high temporal resolution to monitor fire progression every 15 minutes. A grid-growing clustering algorithm was employed to classify polygons into individual fire incidents, enabling a quantitative analysis of fire behavior.
- Fire Modeling with WFA-e: Fire simulations were conducted using WFA-e, incorporating various fire spread models, including Rothermel’s surface and crown fire spread models. Fuel type, topography, and weather data were integrated to run simulations.
- Statistical Analysis: The accuracy of the fire spread models was assessed using error metrics such as ROS residuals, mean absolute error (MAE), mean bias error (MBE), and mean absolute percentage error (MAPE).
- Environmental Factors: The study revealed that the accuracy of fire spread predictions was influenced by environmental variables such as wind speed and fuel moisture content (both live and dead). Low wind speeds and high fuel moisture levels tended to lead to underestimations of ROS, while high wind speeds resulted in overestimations.
- Fuel Types: Different fuel types played a significant role in the accuracy of predictions. Models performed relatively well for shrub, grass, and grass-shrub fuel types, while they consistently underpredicted ROS for timber fuel types.
- Overall Model Accuracy: The models had an average MAPE of 47% for automatic fire simulations, with better performance in shrub, grass, and grass-shrub fuel types. Timber fuel types exhibited the highest MAPE (approximately 67%).
The study found that the model errors and biases were reasonable for simulations performed automatically. It identified environmental variables that might bias ROS predictions, particularly in timber areas where some fuel models might underestimate ROS. Overall, the performance of fire spread models for California aligns with studies developed in other regions, and the models are deemed accurate enough to be used in real-time to assess initial attack fires.
Next Steps from the Research
The study highlighted challenges related to pyroconvection, local wind fields, and the estimation of ROS in timber areas. It recommended the development of improved fire spread models to address these challenges and enhance prediction accuracy.
The study found that while current fire spread models have limitations and biases, they are accurate enough to be used in real-time operational settings, particularly with the capability for manual adjustments and calibration. However, there is a need for ongoing improvements, especially for modeling fire spread in timber areas, predicting crown fire behavior, and considering the effects of pyroconvection. This research contributes valuable insights to wildfire prediction and management, emphasizing the importance of continuously refining and enhancing predictive models in the face of growing wildfire threats.
The research underscores the importance of wildfire simulators in supporting planning and incident analysis in real-time, despite the potential uncertainties derived from input data quality and model inaccuracies. The study additionally provides insights into the performance of fire spread models in California, offering a foundation for understanding and potentially improving upon current operational models in the future.
Learn more about how this science is put into practice.
