As the fire season nears in certain areas in the US, a team of scientists utilized the power of artificial intelligence to predict wildfires. Their model could map the fuel moisture in several states to warn authorities of prone areas.
The innovative technique to predict upcoming wildfires was developed by scientists at Stanford University, a private research university in the US. They created a deep-learning model capable of measuring the fuel moisture levels in certain states. Integrated with satellite data, the model could determine areas where wildfires might likely occur. This method would be significant in the current pandemic. Their findings were described in the August 2020 issue of Remote Sensing of Environment.
COVID-19 and Wildfires
In territories where wildfires are common during the fire season, governments tend to use controlled fires or the intentional and regulated process of burning combustible materials in the forest to prevent wildfires from going out of control. Because controlled burning reduces the number of combustible materials, wildfires are unlikely to spread and lay waste in forests. It also decreases the chances of wildfires reaching human communities, but performing controlled burning is no easy task as it requires a specialized team.
According to the US Forest Service, an arm of the US Department of Agriculture, controlled fires are one of the main reasons why periodic wildfires do not happen. The method contributes as well to the managing of the life cycle of various vegetation. By clearing overly populated plant species, controlled burning allows other plant species to grow and thrive. However, controlled fires are not applicable all the time because several factors must be considered to save resources. If controlled burning is conducted when factors are against it, effort, time, and resources are wasted.
In the ongoing COVID-19 pandemic, there is major concern about wildfires in the upcoming fire season, particularly in the US. Shortages in resources and personnel can limit the ability of the government to send people to conduct controlled burning. If neglected, wildfires can result in disaster and mass evacuation. The latter can be a way for the virus to spread easily. As such, a precision strike is required to avoid wasting resources.
AI-Powered Model to Detect Areas Prone to Wildfires
At Stanford University, scientists created an artificial intelligence (AI) program to determine areas vulnerable to wildfires. Via deep learning, the AI could measure fuel moisture levels or the amount of water in vegetation. Low moisture levels would tag a location susceptible to wildfires, especially in the fire season. Their approach could assist specialists in conducting precise strikes of preventive measures. Although the model still requires further testing, its preliminary performance has already shown patterns not visible before.
"One of our big breakthroughs was to look at a newer set of satellites that are using much longer wavelengths, which allows the observations to be sensitive to water much deeper into the forest canopy and be directly representative of the fuel moisture content," said Alexandra Konings, the senior author of the study and ecohydrologist at Stanford.
In the study, the team explained that fire agencies normally calculate the quantity of flammable vegetation in an area from a small number of trees. The most common method used is to chop and weigh tree branches, dry them in an oven, and weigh them once more. The technique was deemed laborious and could not be applied to all cases. This would contribute to blind spots for fire agencies that might affect their decision-making in prioritizing preventive measures. As an improvement, the US Forest Service established a database of fuel moisture in different sites. But fuel moisture levels in plants could vary from one plant to another and one ecosystem to another.
The deep-learning model developed in this study was based on a recurrent neural network that could recognize patterns in available data. The model was trained by scientists using the data from the National Fuel Moisture Database. Its analysis of patterns would include estimates from two types of measurements collected by satellites. One measurement type was the visible light bouncing off Earth and the other type was the synthetic aperture radar, which measures the return of microwave radar signals.
For the training and validation, the model was presented with three years of data for 239 sites in the 12 states in the Western US. The data was dated from 2015 to 2018. Scientists examined the fuel moisture predictions in six common land cover types, such as broadleaf deciduous forests, grasslands, and shrublands. The most accurate model reading that matched closely with the database was in shrublands. Both the database and the AI model indicated the abundance of aromatic herbs in shrublands.
Scientists agreed that the prediction is correct. First, aromatic herbs tend to grow quickly in favorable conditions. If their fuel moisture levels are low, they can burn easily and spread the fires in a large area. Second, the majority of shrublands are often dominated by aromatic herbs. In the Western US, up to 45% of the area is occupied by shrublands.
According to the Congressional Research Service, a public policy research institute of the US Congress, an estimated 58,100 wildfires were detected in the US in 2018. It corresponded to a total of 8.8 million acres burned by the fires in the same year. Both the instances of wildfires and total acres burned were lower compared to 2017, which recorded 71,500 wildfires and 10 million acres burned. However, 2015 had the greatest impact in the history of wildfires since 1960, wherein 68,200 wildfires occurred and 10.13 million acres of land were razed. Still, out of 1.4 million wildfires since 2000, only 189 of those exceeded 100,000 acres and 13 exceeded 500,000 acres.
Right now, the model only offers historical information in the fuel moisture from 2016 to 2019. But once it is refined, its latest feeds will be directed to an interactive map that fire agencies may use to prioritize fire preventive measures. If the model performs well, fire agencies will no longer have to conduct the random fieldwork to assess areas for wildfires. They only have to do it to confirm the model's findings.