Forecasting not only the likelihood but also intensity of wildfires is important for both industry, safety, and better understanding climate change. More recently, we have seen the effects of smoke from wildfires and how deleterious effects on health can occur sometimes hundreds of miles away from where the fires are.
Satellite systems are vital for observation of developing fires, where they can be used to monitor conditions which can then feed information to forecast the likelihood of fires using developed models. Techniques integrating artificial intelligence and, in some cases, combining wildfire and smoke forecasting are likely to be increasingly important in an era where large-scale wildfires are becoming the norm.
Using remote sensing data to forecast wildfires
NASA’s Global Fire Weather Database (GFWED), which provides wind, temperature, and humidity data that can then be used with GIS or remote sensing software to forecast where fires could begin, continues to be an important global database used for forecasting wildfires.
GFWED uses satellite-based precipitation measurements made in near real-time that are applied to forecast the probability that a wildfire could begin in a given area. Effectively, the model provides a type of fire score that shows areas where wildfires are more likely.
The model uses past patterns of fire events to help determine likely areas where wildfires are more likely to easily spread. For example, it is not just how dry a place is but also the fact that wind speeds may have picked up that could affect one given area. Combing different factors that lead to developing wildfires is how the tool provides a better regional understanding of fire occurrence.
Using artificial intelligence to forecast and map wildfire probability
Machine learning and artificial intelligence techniques, including random forest and genetic algorithms, have been used to estimate the likelihood of fires using such imagery as Moderate Resolution Imaging Spectroradiometer (MODIS) data.
In one type of model, factors such as humidity, wind speed, rainfall, elevation, slope, and normalized difference moisture index (NDMI) are used to predict where wildfires could start. Techniques such as random forest appear to be among the most accurate, with accuracy near 88%.
There are, however, other approaches that have been developed that utilize both human and natural factors in assessing fire risk.
For instance, using MODIS data, a learning algorithm using expectation-maximum methods was used to evaluate a series of inputs from different regions. Bayesian networks and GIS were used to evaluate regional factors of different inputs influencing fires from historical data. This was applied in Eswatini and it was found that both land tenure and land cover conditions were shown to be major reasons for fire, where the model was able to have accuracy of over 93% in forecasting fires.
Satellite data can be used to train the model which can then use the previous data to forecast future conditions that might be similar to past occurrences.
Increasingly, newer models are looking at spatial resolution, particularly localized factors that determine how a fire spreads. These recent models applied localized variables to enable even more precise understanding or forecasting of wildfires.
In this case, a Location-aware Adaptive Normalization layer (LOAN) is used where dynamic variables (e.g., weather conditions) are separated from more static variables (e.g., topography); this helps the deep learning convolutional neural network better forecast how wildfires evolve in a given and more precise area.
Spatio-temporal 2D/3D convolutional neural networks allow spatial and temporal variables to be split and applied for forecasting in very localized areas. Such results demonstrate that deep learning is now beginning to become a promising way in which wildfires can be forecasted. Deep learning approaches are much more data intensive and this means this will require much more precise monitoring techniques.
Data may need to better capture both satellite-based and local ground-based observations if deep learning models are to have high levels of accuracy in forecasting likely pathways of wildfire spread.
Understanding Human Factors and Using Remote Sensing and GIS for Fire Risk
Other methods have tried to combine GIS, remote sensing and interviews with local experts to better understand fires. In these cases, it has been observed in some regions, such as the Mediterranean area, a high percentage of fires are started by human factors.
Understanding the right conditions from earth observation techniques and GIS could be improved by also knowing how these relate to human factors, such as reasons why deliberate actions have been done to lead to large-scale fires. This helps to produce an accurate understanding of human-environment risk factors in fire occurrence.
Direction of Software Development
Many national forest services use standardized models to help predict and manage wildfires. Software tends to use fuel and moisture conditions, simulating surface and fire rate and spread.
In the United States, the standard U.S. Department of Agriculture Forest Service model is BehavePlus, a simulation based approach used to forecasting wildfire spread. Increasingly, researchers have been arguing that AI-based forecasting tools may provide better results.
Factors that incorporate natural and human factors are likely to be how tools continue to evolve. In particular, machine learning approaches can also be flexible in how incorporating different factors based on seasonal variation could help them outperform some existing statistical or simulation-based approaches.
Forecasting the spread of wildfire smoke
Scientist are not just concerned about forecasting wildfire spread. Greater attention has now been given to forecasting the spread of smoke, given that smoke can travel much further than fires and can affect the health of millions of people through the dispersion of fine particle matter.
The High-Resolution Rapid Refresh Smoke (HRRR-Smoke) is an experimental wildfire smoke prediction model that is an extension of NOAA’s existing HRRR weather model predicting rain, wind, and thunderstorms. This model incorporates real-time data from the Joint Polar Satellite System’s Suomi-NPP and NOAA-20 polar satellites, in addition to data from NASA’s Terra and Aqua satellites.
In the United States, the National Weather Service uses the HYSPLIT-based system to forecast smoke drift and air quality. The model is essentially a dispersion forecasting tool that looks at transport, dispersion, chemical transformation, and deposition in simulating likely areas smoke could travel.
Recent work has integrated this framework with a fire spread model, called ELMFIRE, that forecasts climate change impact studies as well as recent events. By combining fire forecasting techniques and inputs with plume and particle matter spread, the approach of integrating these models appears to yield very accurate results, comparable to HYPSPLIT, without the need for having a lot of input data.
More specifically smoke spread can be forecasted at high accuracy levels by focusing on modeling where several large or key fires spread rather than focusing on a capturing a large number of current fires. Key fires can help outline important inputs in the atmosphere and weather conditions that drive smoke spread.
This demonstrates how health officials as well as fire managers may be able to quickly obtain an idea of what areas are likely to face air quality issues as wildfires develop. Rather than waiting to get all needed wildfire data, officials may simply focus on some of the key fires to estimate which areas are likely to be affected by smoke.
Future directions in modeling and mapping wildfires
Wildfire modeling is now beginning to expand well beyond simulation and statistical methods by incorporating new forms of artificial intelligence techniques. Many of the older models are very accurate and useful in their forecasting capabilities, but deep learning methods can potentially help fill uncertainty gaps.
Moreover, with improved monitoring that combines satellite-based and ground-based methods, more localized models can be used to better forecast overall likely fire behavior.
This, however, suggests we will need far better monitoring that integrates the various levels of observation in order to use these more data-intensive deep learning models, which may remain a challenge for some time to come in many parts of the world. In other words, older fire models are likely to remains some of the best approaches until we can improve our data capture in many remote areas.
With health hazards of individuals far away from fires now also a major concern, future methods will likely integrate fire and smoke models more closely. Rather than focusing on areas immediately impacted by fires, models will look at a number of surrounding fires that may together shape air quality at continental scales.
Earth observation techniques have been used to forecast fires since at least the 1960s and the early days of satellite-based remote sensing. What is different now is that models at global and regional scale are being developed that have been increasing in accuracy. With climate change now becoming a real risk for fires, new techniques are also incorporating land-based data along with temperature, precipitation and other weather factors in developing more accurate models.
 For more on fire forecasting using NASA’s GFWED, see: https://earthobservatory.nasa.gov/images/92367/forecasting-fire.
 For a recent study on using machine learning and artificial intelligence techniques on wildfires, see: Abdollahi A and Pradhan B (2023) Explainable artificial intelligence (XAI) for interpreting the contributing factors feed into the wildfire susceptibility prediction model. Science of The Total Environment 879: 163004. DOI: 10.1016/j.scitotenv.2023.163004.
 For more on using MODIS and learning algorithms for forecasting fires, see: Dlamini, W. M. (2011). Application of Bayesian networks for fire risk mapping using GIS and remote sensing data. GeoJournal,76(3), 283–296. https://doi.org/10.1007/s10708-010-9362-x.
 Eddin MHS, Roscher R and Gall J (2023) Location-aware Adaptive Normalization: A Deep Learning Approach For Wildfire Danger Forecasting. IEEE Transactions on Geoscience and Remote Sensing: 1–1. DOI: 10.1109/TGRS.2023.3285401.
 For more on understanding human factors and using remote sensing and GIS for fire risk understanding, see: Leone, V., Lovreglio, R., Martín, M. P., Martínez, J., & Vilar, L. (2009). Human Factors of Fire Occurrence in the Mediterranean. In E. Chuvieco (Ed.), Earth Observation of Wildland Fires in Mediterranean Ecosystems(pp. 149–170). Berlin, Heidelberg: Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-01754-4_11.
 Wang W, Zhao F, Wang Y, et al. (2023) Seasonal differences in the spatial patterns of wildfire drivers and susceptibility in the southwest mountains of China. Science of The Total Environment 869: 161782. DOI: 10.1016/j.scitotenv.2023.161782.
 Stein AF, Draxler RR, Rolph GD, et al. (2015) NOAA’s HYSPLIT Atmospheric Transport and Dispersion Modeling System. Bulletin of the American Meteorological Society 96(12): 2059–2077. DOI: 10.1175/BAMS-D-14-00110.1.
 Melecio-Vázquez D, Lautenberger C, Hsieh H, et al. (2023) A Coupled Wildfire-Emission and Dispersion Framework for Probabilistic PM2.5 Estimation. Fire 6(6): 220. DOI: 10.3390/fire6060220.
This article was originally written on July 22, 2018 and has since been updated.