· 11 min read
As machine learning has become more accessible and models have become easier to develop, it has become a go-to modeling technique for predicting hazards related to wildfires and as a tool for developing mitigation solutions. Machine learning is faster and more flexible than simulation and easier to run on a large landscape with relatively limited computing resources. As it predicts wildfire characteristics across a landscape, machine learning can simultaneously generate valuable insights about larger phenomena related to wildfires that can only be observed at a landscape scale.
In this article, I will share some of the results and insights from FireCasterAI which is a web app built around a deep-learning model for predicting and quantifying wildfire hazard and risk characteristics in the western United States at both weather and climate timescales. All the maps in this article were generated using FireCasterAI. The climate-related charts are from ClimateVision, a simplified climate data platform.
Landscape Details and Fire Weather Index
The primary inputs to the machine learning model are the details of the landscape and the local weather conditions during a wildfire event.
Figure 1 below shows landscape details of the twelve western states in the US covered by the model. Large swaths of the landscape consist of grasslands and shrublands, particularly in the southwest and areas farther from the coast. Oregon includes two distinct types of landscapes: high canopy cover in western Oregon, and mostly grasslands and shrublands in southeastern Oregon. Similarly, eastern Texas has a high canopy cover while western Texas is mostly grassland and shrubland. Montana is divided roughly into two parts as well, with the majority of forestlands occurring in the western part while grasslands and shrublands dominate the eastern side. This pattern of vegetation greatly influences the potential for wildfires on these landscapes.
Figure 1. Landscape details (Source: FireCasterAI)
The Fire Weather Index (FWI) accounts for the effects of fuel moisture and weather conditions on wildfire behavior. FWI is widely used as an indicator of fire danger with higher values indicating higher danger. It is calculated using hourly and daily measurements or forecasts of temperature, relative humidity, precipitation, and wind at a given location. Figure 2 shows the distribution of average FWI across the western states for the month of June in 2021-23. The regions with high FWI values are also roughly the landscapes dominated by grasslands and shrublands.
Figure 2. Fire Weather Index for June 2021-23 (Source: FireCasterAI)
Wildfire hazard patterns
The map in Figure 3 shows the distribution of the Wildfire Hazard Index (WHI) across the twelve western states as computed by FireCasterAI’s machine learning models. The map essentially depicts the forecasted distribution of wildfire hazard (or potential) on a specific day in June this year, with the darker orange indicating higher WHI values.
This pattern of wildfire hazard closely follows the FWI distribution as well as the vegetation. Wildland areas dominated by grasslands and shrublands typically have higher WHI values since these types of vegetation can dry more easily than forestlands and accumulate dry fuel that is ready to burn in the fire season. The wildfire event then simply depends on whether there is an ignition event or not. In states such as Oregon, Montana and Texas, the vegetation varies considerably between different parts of the state and the wildfire hazard essentially follows this same pattern.
Figure 3. Wildfire Hazard Index for June 8, 2025 (Source: FireCasterAI)
Urban wildfire exposure
Urban regions are essentially small and large islands in a vast sea of vegetation, as Figure 1 shows. This means that nearly every urban area has some exposure to wildfires spreading into that area from nearby wildlands. This exposure varies widely and is a function of the wildfire potential in the nearby wildlands at any given time as well as the proximity of the urban area to those wildlands. As examples, the maps in Figures 4-6 show the Urban Hazard Index values (indicated by shades of red, with darker shades indicating higher values) in three major urban areas and the Wildfire Hazard Index values (indicated by shades of orange) in the surrounding wildlands.
Albuquerque, New Mexico, is surrounded entirely by grasslands/shrublands that have fairly high wildfire hazard index values, and consequently the urban area has high urban hazard index values that indicate an elevated risk of exposure to wildfires from nearby wildlands. The Portland-Vancouver metropolitan area with its high canopy cover has relatively low urban hazard values while smaller towns in eastern Oregon have higher exposure to wildfires from the surrounding grasslands and shrublands. Similarly, Dallas and Austin in the state of Texas have fairly low urban hazard values, while smaller towns in western Texas have higher urban hazard values because of the surrounding vegetation.
Figure 4. Urban and Wildfire Hazard Indices in and around Albuquerque, New Mexico - May 25, 2025 (Source: FireCasterAI)
Figure 5. Urban and Wildfire Hazard Indices in and around Portland, Oregon - June 8, 2025 (Source: FireCasterAI)
Figure 6. Urban and Wildfire Hazard Indices in and around Dallas, Texas - May 25, 2025 (Source: FireCasterAI)
The wildfire-climate connection
As the climate changes in the coming decades, the weather variables that are closely related to wildfires will also change. The primary weather variable here is temperature, but precipitation, relative humidity and wind are also key contributors to the Fire Weather Index. Figures 7 and 8 show the average Wildfire Hazard Index values for the month of June in the recent past (2021-2023) and at mid century in the year 2050 as computed by FireCasterAI’s machine learning models.
Figure 7. Wildfire Hazard Index - June 2021-2023 (Source: FireCasterAI)
Figure 8. Wildfire Hazard Index - June 2050 (Source: FireCasterAI)
The patterns in the two maps look remarkably similar, even though climate is expected to be different in 2050 based on current emission trajectories. A closer examination of the data underlying these maps shows in fact that there are more wildland locations in 2050 with higher hazard values than in 2021-2023, but the increase is not dramatic. The number of locations with the highest hazard index values is projected to increase by less than 7%.
In order to understand this better, let us look at one specific location where the wildfire hazard does not seem to have worsened in 2050: Portland, Oregon. Figures 9-11 (generated by the ClimateVision platform) show the evolution of temperature, precipitation, and relative humidity at this location between 2020 and 2100 under two emission scenarios according to the CMIP5 global climate models downscaled to North America. Of these three variables, only the temperature shows a consistent increasing pattern, but the actual increase by 2050 (relative to 2020) is no more than 1.5oC in the worst case.
Figure 9. Temperature trend in Portland, Oregon, under two emission scenarios - 2020 to 2100 (Source: ClimateVision)
Figure 10. Precipitation trend in Portland, Oregon, under two emission scenarios - 2020 to 2100 (Source: ClimateVision)
Figure 11. Relative humidity trend in Portland, Oregon, under two emission scenarios - 2020 to 2100 (Source: ClimateVision)
The average Fire Weather Index in Portland for the month of June is projected to increase from 7.6 in 2021-23 to 10.4 in 2050 (both of which are considered very low), which shows that the average warming trend by itself is not a significant factor for wildfire hazard at this location. In the Angeles National Forest north of Altadena (which was one of the sites of the costly 2025 Los Angeles wildfires), the average FWI is projected to increase very marginally from 60.4 in June 2021-23 to 61.4 in 2050 – confirming again that the general warming trend is not the driving force behind the increasingly larger and more frequent wildfires we are starting to see.
Is there another explanation? One possibility is the phenomenon known as hydroclimate whiplash. In the Los Angeles region, the 2022-23 and 2023-24 water years were both very wet which led to a buildup of vegetation in the nearby wildlands. This was followed by the anomalously hot summer-fall period of 2024, which dried out the built up vegetation and likely created the conditions for the large wildfires that occurred in January 2025. In other words, excessive wet periods followed by very hot and dry periods over a 2-3 year time frame can significantly increase wildfire risk, especially when there are also strong winds like the Santa Anas. So climate change does play a role here but more through short-term swings between extremes rather than through any long-term average trends – at least until about 2050.
Causes of wildfires
Figure 12 shows the distribution of the likely causes of wildfires that might occur at the peak of the fire season across the western states, as predicted by FireCasterAI’s machine learning models. Human causes – which range from recreational activities to sparking power lines and equipment – dominate the west coast and many of the large population centers. Lightning – the only natural cause of wildfires – is the predominant cause across sparsely populated areas, higher elevations and wildlands that are not in proximity of large urban areas.
Figure 13 depicts the distribution of the Lightning-induced Fire Risk Index, which combines the Wildfire Hazard Index with the average power density from lightning at every location. This risk distribution aligns closely with the causes of wildfires and is generally more intense at higher elevations and away from the coast.
Figure 12. Causes of wildfires - July 2021-2023 (Source: FireCasterAI)
Figure 13. Lightning-induced Fire Risk Index - July 2021-2023 (Source: FireCasterAI)
Wildfire mitigation and urban exposure reduction
The results from the machine learning models can be used as part of a whole systems approach that yields a hierarchy of solutions to mitigate damage from wildfires. As alluded to earlier and illustrated by Figure 1, major urban areas are islands in a much larger sea of wildland vegetation and it is this vegetation that provides the initial fuel for wildfires to start and spread. One of the key insights generated by the machine learning models is how the wildfire hazard posed by wildland locations turn into urban fire hazards at nearby urban areas. Another insight is about the type of vegetation: grasslands and shrublands tend to dry easily and pose a higher risk for nearby urban areas than forestlands.
Recent experience confirms some of this. Two of the largest urban fires in the last few years – the 2025 Los Angeles fires and the 2023 Maui fires – started as wildland fires in grassland or shrubland. Safe removal of dry fuel in these landscapes through techniques like prescribed burns will be critical for protecting nearby urban areas.
Landscapes with sufficient tree canopy cover appear to have lower wildfire hazard values than grassland or shrubland according to the machine learning models. This pattern is obviously something that the models have learned from training on historical data. However, if a surface fire reaches the crowns of trees, fires can become intense and spread rapidly, so using mechanical thinning to keep the canopy density under a safe threshold and remove smaller trees that can act as ladder fuels will be important for maintaining these lower hazard levels.
Closer to the urban areas, it is relatively easy to identify the exposure to wildfires based on the vegetation, topography and climate/weather characteristics of nearby wildlands using methods like machine learning as we have shown here. Armed with this information, urban areas can be made more wildfire resilient using a number of techniques. Buffer zones strategically located at the periphery of urban communities could prevent fire spread into the communities and could be a cost effective way of protecting entire communities. Deeper into the communities, defensible spaces around buildings and fire-resistant structures would be the last line of protection especially where the urban hazard levels are high due to the surrounding wildlands.
illuminem Voices is a democratic space presenting the thoughts and opinions of leading Sustainability & Energy writers, their opinions do not necessarily represent those of illuminem.
