Limber, R., Hargrove, W. W., Hoffman, F. M., & Kumar, J. (2024). Forecast of Wildfire Potential Across California USA Using a Transformer. Proceedings of the 2nd International Workshop on Big Data Analytics with Artificial Intelligence for Climate Change at IEEE Big Data 2024, IEEE, Washington, D.C., USA, pp. 4324–4332. DOI: https://doi.org/10.1109/bigdata62323.2024.10825778

Forecast of Wildfire Potential Across California USA Using a Transformer

Transformers can effectively emulate the US Forest Service’s Wildland Fire Potential Index maps using fewer, easier to obtain input variables, with improved computational efficiency

The Science

The authors used meteorological variables from Daymet, the normalized difference vegetation index (NDVI) calculated from the Moderate Resolution Imaging Spectroradiometer (MODIS) and the Scott and Burgan fire behavior fuel model from LANDFIRE, as inputs for a transformer with a residual connection to model Wildland Fire Potential Index (WFPI). The area of interest for this analysis was California USA. The US Forest Service’s WFPI maps were used to train the transformer. The transformer was tuned using Bayesian hyperparameter optimization with an architecture designed for time series analysis, whereby the past 30 days of inputs were used to forecast output for the subsequent WFPI value. Put more precisely, given the residual connection, the transformer outputted a bias that shifted “yesterday’s” WFPI, to align with what it perceived to be “today’s” WFPI values, thereby predicting subsequent WFPI values over the target region. This process can be repeated in series to forecast k days into the future. In this analysis k was restricted to seven to align with the forecast window of the US Forest Service’s WFPI model.

The Impact

Using open access meteorological variables, NDVI and the Scott and Burgan fire behavior fuel model, the transformer with a residual connection proved to be capable of effectively emulating the US Forest Service’s WFPI maps. Looking at a seven day forecast (equal to the forecast range of the WFPI model) for July of 2023, the transformer emulated maps compared to the WFPI modeled maps have correlations ranging between 0.85 - 0.98. The transformer can be easily parallelized and was able to predict seven day forecasts for each day of 2023 in only 6.5 minutes using two 64 core CPUs.

Summary

Wildfires are a major issue facing the United States. In California alone, wildfires are responsible for billions of dollars in damages and take lives each year. Accurately predicting fire danger conditions allows preparation awareness before wildfires start. Transformers are a class of deep learning models designed to identify patterns in sequential datasets. In recent years, transformers have gained popularity through their impressive performance in natural language processing and other applications of signal recognition. This analysis demonstrates the ability of a transformer with a residual connection to forecast fire danger potential over the state of California. WFPI maps collected from the US Geological Survey database from January 1st 2020 to December 31st 2023 were used to tune, train and evaluate the transformer. Meteorological inputs (provided by Daymet daily weather and climatological summaries), the NDVI calculated from MODIS, and the Scott and Burgman fire behavior fuel models (to characterize maps of fuel types), were used as inputs. Our results show that a transformer can effectively emulate the US Forest Service modeled WFPI maps of California USA for four week long forecasts over the month of July, 2023, with correlations ranging from 0.85 – 0.98.