Seasonal Vapor Pressure Deficit Guidance Data & Documentation

The Physical Sciences Lab provides experimental vapor pressure deficit (VPD) guidance to provide information that can be used in fire weather risk scenario development. These forecasts are based on a machine learning approach using a linear inverse model (LIM; Penland and Sardeshmukh 1995). The LIM relies on observed relationships between model variables (Table 1) to infer predictable Earth system processes and generate probabilistic forecasts.

Vapor Pressure Deficit

Vapor pressure deficit (VPD) is defined as the difference between the saturation vapor pressure e_s(T) and actual vapor pressure e_a(Td), which is calculated using 2-meter temperature T and 2-meter dewpoint temperature Td from the ERA5 reanalysis. VPD is a measure of near-surface atmospheric moisture demand that is strongly and positively correlated with the annual mean area burned by wildland fires in the western United States (Williams et al. 2014; Abatzoglou and Williams 2016; Higuera and Abatzoglou 2021). Higher VPD corresponds to greater area burned as it reflects drier conditions that are conducive to wildfires.

Method

The LIM is trained using monthly differences from average with a three-month running mean applied, and the covariance statistics between VPD, sea surface temperatures, and soil moisture are used to train the LIM using the years 1958-2022. For assessing forecast skill relative to a more current baseline, the resultant forecast anomalies are re-centered on the current 30-year based period used by the Climate Prediction Center, which is 1991-2020. Sea surface temperature and soil moisture are used because they are key predictors on seasonal timescales that have strong relationships with VPD. Similar to output from numerical forecast models, for each initialization and lead time, the LIM generates forecasts of VPD and SST by propagating the initial conditions forward in time. Near-real-time forecasts are created using the most recent month's ERA5 data (Hersbach et al. 2020).

Deterministic forecasts are generated for each lead time and initialization date using the forecast operator, and subsequently, probabilistic forecasts are generated by creating a forecast PDF centered around the deterministic forecast. The spread of the PDF is lead-dependent — not state-dependent — and is calculated at each lead time using the forecast error covariance following Penland and Sardeshmukh 1995.

Data Source

All variables used for training and initialization of forecasts are from ERA5 data page.

CONUS VPD LIM

One important way to check how well a model performs is to have it "predict" things that we already know — this is referred to as hindcasting. The year-round western US (see Figure 1) LIM hindcast skill (shown in Figure 2 below) uses a method for assessing skill called anomaly correlation coefficient (ACC), which measures the strength of the shared variance between observations and the forecast. LIM VPD skill is statistically significant and frequently exceeds 0.5, a commonly-accepted value indicating that on average, the model is skillful enough to be useful. Furthermore, the LIM VPD skill outperforms a persistence forecast (Figure 2b) which can be a competitive forecast for temperature-related quantities.

Another measure of forecast skill is reliability (Figure 3 below), which assesses how the probabilities from a forecast align with the observed frequency of a certain outcome, such as VPD being above or below the median value. If forecasts are perfectly reliable, then the line plotted will follow the one-to-one line (dashed line). Hence, it can be seen that LIM VPD forecasts at a three-month lead time, as one example, are close to the one-to-one line and therefore are highly reliable. This gives us confidence in the probabilities that are displayed in the forecasts.

Table 1: Description of variables used to construct the VPD LIM including sea surface temperature (SST), soil moisture (SM) and vapor pressure deficit (VPD). ERA5 Reanalysis is used for all variables.