Using deep learning to improve multi-day forecasts of extreme fire weather and behaviour throughout the United States

The ignition, spread, and severity of wildfires are driven largely by weather conditions (Jain et al. 2020: https://doi.org/10.1139/er-2020-0019; Liu et al. 2013: https://doi.org/10.1371/journal.pone.0055618). The main tool for weather prediction across the globe is a set of physical, coupled atmosphere/ocean models, called numerical weather prediction (NWP). Despite rapid improvements in the last few decades, NWP alone is not sufficient for wildfire prediction, because it does not resolve every process related to wildfire. One solution is to post-process NWP with statistical models, which correct the NWP model towards better resolving processes related to the phenomenon of interest (here, wildfire). This post-processing is called model-output statistics (MOS) and typically involves linear regression. However, recent work has advanced MOS by incorporating more powerful statistical models from deep learning (DL). We use DL to predict extreme fire weather and behaviour at multi-day lead times throughout the United States. For fire weather, we have trained U-nets -- a type of deep neural network -- to predict at lead times of 3-240 hours over the United States. The output (target) variables are seven indices from the Canadian Fire Weather Index System (CFWIS), computed from the ECMWF Reanalysis version 5 (ERA5). These seven indices include the fine-fuel moisture code (FFMC), initial-spread index (ISI), overall fire-weather index (FWI), etc. Meanwhile, the input (predictor) variables come from five sources. The first is a forecast time series of atmospheric state variables (height, temperature, humidity, and wind) from the Global Forecast System (GFS) NWP model. The second is a forecast time series of surface and subsurface moisture (soil moisture, accumulated precipitation, and snow depth) from the GFS. The third is a set of constant fields (terrain height/slope/aspect, land-sea mask, etc.) describing the underlying terrain. The fourth is a lagged time series of CFWIS over the past several days, i.e., past target values. The fifth is a forecast time series of CFWIS indices, computed by applying the CFWIS functions directly to GFS-forecast weather variables. These are the uncorrected (GFS-only) CFWIS forecasts, to be corrected by the U-net. For fire behaviour, we have trained random forests -- ensembles of decision trees -- to predict fire radiative power (FRP) at lead times of 1-48 hours over the United States. The labels (correct answers) for FRP are obtained from the Regional ABI and VIIRS Emissions (RAVE) merged satellite product. Predictors for the random forest include the first three sources listed for the U-net above, plus a lagged time series of FRP over the past 24 hours, i.e., past target values. Both models -- the U-net for fire weather and the random forest for fire behaviour -- are trained with built-in uncertainty quantification. Thus, at every lead time and grid point, both models provide an expected value and an estimate of their own uncertainty. We will present objective evaluation results (for both the mean forecast and uncertainty) and explainable artificial intelligence (XAI) to understand what the models have learned, e.g., which spatiotemporal weather patterns in a given area are most conducive to extreme fire weather/behaviour.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.