Forecasting Extreme Fire Weather across the United States: A Deep Dive with Deep Learning, Uncertainty Quantification, and Explainable AI

The ignition, spread, and intensity 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). Fire weather is usually quantified by indices derived from basic weather variables, such as 2-m temperature/dewpoint, 10-m wind, and precipitation. Hence, the main tool for large-scale prediction of fire weather is numerical weather prediction (NWP). However, NWP models are not specifically tuned for fire weather, leaving room for improvement via post-processing. Also, most NWP models do not provide adequate uncertainty quantification – single-member models provide only one deterministic forecasts, and ensemble models are notoriously underdispersive. Thus, we use machine learning with uncertainty quantification (ML-UQ) to post-process NWP forecasts of fire weather.

Specifically, we post-process Global Forecast System (GFS) forecasts of fire weather at short- to long-range lead times (1, 2, ..., 14 days) across the entire United States, including Alaska and Hawaii. We predict the seven variables making up the widely used Canadian Fire Weather Index System (CFWIS), which quantify fuel moisture in light to heavy fuels, potential fire spread, and potential fire intensity. The seven target variables are the fine-fuel moisture code (FFMC), duff moisture code (DMC), drought code (DC), initial-spread index (ISI), buildup index (BUI), fire-weather index (FWI), and daily severity rating (DSR). Truth labels are computed by applying the CFWIS calculations, codified in the fwi.fbp software package (Wang et al. 2019: https://r-forge.r-project.org/projects/fwi-fbp/), to weather fields from the ECMWF Reanalysis version 5 (ERA5).

Our model is a deep neural network with a modified U-net++ architecture (Zhou et al. 2019: https://doi.org/10.1109/TMI.2019.2959609), taking predictors from five data sources. The first is a GFS-forecast time series of atmospheric state variables (temperature, humidity, wind, and pressure/height) at levels including the surface, 900 hPa, and 700 hPa. The second is a GFS-forecast time series of surface and subsurface moisture – namely soil moisture, accumulated precipitation, and snow depth. The third is a set of time-invariant fields related to terrain – including orographic height, slope, aspect, land/sea mask, etc. The fourth is a lagged time series of truth variables (ERA5-derived CFWIS indices) over the past several days. The fifth and last data source is a forecast time series of the target variables (CFWIS indices), computed by applying fwi.fbp directly to GFS weather fields. This data source also constitutes the baseline (raw fire-weather forecasts), which we compare with forecasts from the U-net++ model. The single U-net++ produces an ensemble forecast of every target variable at every lead time at every grid point.

To suit this particular application, we modify the basic U-net++ methodology as follows. First, we use a custom loss function called the dual-weighted mean squared error (DWMSE) skill score. The “dual-weighted” part emphasizes extreme values of all fire-weather indices, and the baseline for the skill score is the raw GFS forecast valid at the same time. Thus, our loss function emphasizes improvement over the raw GFS, especially for extremes. Second, because two of the target variables are uniquely related to each other – namely, DSR = 0.0272 * FWI^1.77 – we enforce this relationship in the model architecture. Third, we use residual prediction, where the U-net++ is tasked only with predicting the difference between the current truth fields (CFWIS indices at forecast-issue time) and those at the valid time (CFWIS indices 1-14 days ahead). Residual prediction improves skill especially for slow-changing indices, such as the DMC, DC, and BUI – for which there is usually little change over 1-14 days, making a persistence forecast already quite skillful. Fourth, we use a learning curriculum to train a single U-net++ model for all 14 lead times. Early training epochs focus on the 1-day lead time (coarse tuning); later epochs focus on the longer lead times (fine tuning); and predictors vary according to the U-net++ lead time. For example, when the U-net++ lead time is 1 day, predictors do not include GFS forecasts beyond 2 days; but when the U-net++ lead time is 5 days, predictors may include GFS forecasts out to 10 days, as more temporal context is needed. Fifth, our ML-UQ approach is a hybrid between two methods – the multi-model ensemble (MME) and continuous ranked probability score (CRPS) – allowing the U-net++ to capture both aleatory and epistemic uncertainty. The MME approach is achieved by training multiple U-net++ models with different random seeds; the CRPS approach is achieved by training each model to produce its own ensemble, constrained with a loss function that combines the CRPS and DWMSE skill score, the former encouraging good ensemble calibration.

We will demonstrate that the deterministic version of the U-net++ (ensemble mean only) outperforms the GFS for all learning tasks – i.e., all pairs of lead time and CFWIS index. Furthermore, we will demonstrate the U-net++ provides well calibrated uncertainty for most tasks. Lastly, we will compare explainable artificial intelligence (XAI) approach for distilling insight from the U-net++, namely for quantifying the most important synoptic weather patterns for extreme fire weather in different regions of the country.