Predicting Wildfire Fuel Moisture from Atmospheric Data Using Machine Learning

Increasing risk from wildfires is a well-known consequence of global climate change, with both their frequency and intensity having increased over the last 25 years, and further increases expected due to continued warming and disruption of weather patterns. In addition to directly threatening human life and the property, humans may also be exposed to dangerous emissions, debris flows, water contamination and other impacts. In order to better manage and combat future wildfire danger, reliable predictions are needed so that well-informed planning for infrastructure investment, public safety, and emergency response can be accomplished. This work develops a framework to examine predictions of fuel moisture (FM), a key driver of wildfire risk, under future climate scenarios. A machine learning (ML) approach has been developed, based on past observations of fuel moisture in relation to historical atmospheric data. The trained ML model when applied on the atmospheric data preceding a desired time can be used to construct a map of FM at the desired time. The map of FM can subsequently help assess the risk of wildfire.

In order to assess fuel moisture fuel classes are grouped according to the timescale at which they equilibrate with ambient moisture. This timescale is related to the size of the fuel element, e.g.1-hour fuel corresponds to fuel elements less than 0.25” in diameter, 10-hour fuel to elements between 0.25” and 1” and so on. Here, we concentrate on the 10-hour FM category, since it is the most widely measured and correlates well with the overall wildfire potential. The FM data can be used without modification if using regression ML models or can be sorted into desired number of bins, for example, FM₁, FM₂, ..., FM₁₀, if a classifier algorithm is used. The ML approach is used to find a relationship between FM and the preceding atmospheric conditions (temperature, wind speed, relative humidity, and other parameters) over a specified time interval (e.g. 30 hours) at a specified time frequency (e.g. 3 hours). As a master dataset we use an hourly fuel moisture reanalysis product for California available at 3km resolution. This reanalysis combining observed and modeled data from 2000- 2020 (21 years) available every hour over much of the state of California provides the best estimate of the spatial and temporal fuel moisture variability. A randomly sampled subset in time and space is used for training and testing. The sampling in time helps account for the diurnal, seasonal, and other sources of variability over these 21 years. The sampling in space (grid locations) helps account for the topographical variations. Once an ML model is trained and its accuracy is assessed, it can be applied to an actual wildfire or area of interest.

Here, we assess the application of this approach to historical fires such as Creek fire of 2020, for which the predicted FM map can be compared with FM map from the master data set mentioned above. The trained model will also be applied to future climate simulations to obtain FM maps under future climate scenarios. Future work will pursue increasing the accuracy of the prediction and explore more advanced deep learning (DL) techniques.

Acknowledgements

This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 and was supported by the LLNL-LDRD Program under Project 22-SI-008.