Near-term fire weather forecasting in the Pacific Northwest using 500-hPa map types

Study area

Our study area is composed of the 12 predictive service areas (PSAs) in the Northwest geographic area, which encompasses the states of Washington and Oregon in the northwestern United States. PSAs represent the spatial unit at which the NWCC and other regional coordination centres produce both seasonal and subseasonal fire weather forecasts. In the NWCC, PSA boundaries were drawn to encompass Remote Automatic Weather Stations (RAWSs) with highly correlated measurements of daily minimum relative humidity (Marsha 2014). The study period spans the 20 fire seasons from 2001 to 2020, with earlier years excluded owing to inconsistencies in the RAWS data. We define the fire season to include the months of June through September, which account for 84% of all fires and 91% of the burned area in the region over the study period.

The 12 PSAs in the Northwest are grouped into three broad regions: westside, central and eastside (see Fig. 1). The four westside PSAs (NW01–NW04) include the most densely populated areas in the region as well as the moist coniferous forests of the Olympic Peninsula and the western slopes of the Cascades. NW04, located in southwest Oregon, is more mountainous than the other westside PSAs owing to its inclusion of the northern Klamath Mountains. The three central PSAs (NW05–NW07) lie in the rain shadow of the Cascades and are dominated by dry coniferous forests. The five eastside PSAs (NW08–NW12) are more varied: NW08, NW09 and NW11 are mountainous, whereas NW10 and NW12 include significant areas of shrubland, grassland and cultivated crops.

Fig. 1.

Predictive Service Areas (PSAs) of the Northwest geographic area and land cover classifications from the 2011 National Land Cover Database (NLCD), with key RAWSs used to calculate daily fire danger in each PSA indicated.

The PSAs vary greatly both in terms of number of annual fires and area burned (Fig. 2). The heavily populated westside PSAs of NW01 and NW02 have the smallest annual burn areas despite their moderate ignition counts, indicating that fires that ignite in these areas do not grow beyond a fairly small size. In contrast, average fire sizes are much larger in the eastside PSAs of NW10 and NW12, which experience the largest annual burn areas resulting from relatively few annual fires. These differences in average fire size are likely due to variation in both landcover and wildfire suppression priorities.

Fig. 2.

Average number of fires and area burned per PSA, fire seasons 2001–2020.

Data

In this study, we construct and analyse a longitudinal dataset combining daily wildfire occurrence data with 500-hPa map types and fire danger indices. Data on the occurrence of wildfires are drawn from the Fire Program Analysis fire-occurrence database (Short 2022). Each fire in the dataset includes a point of origin, discovery date and final burned area. Wildfires are spatially matched to PSAs using their point of origin. We further classify fires as ‘large’ based on whether their final size exceeds the operational threshold for significant fires, which is defined as fires that are likely to require the mobilisation of outside resources (Marsha 2014). In the Northwest, significant fires are identified as those exceeding PSA-specific size thresholds (Table 1), which approximately correspond to historical 95th percentile fire sizes within each PSA. Focusing on fires that are relatively large within a given region, in addition to considering total ignition counts, is of operational significance because these large fires require the largest amount of suppression resources (Nagy et al. 2018).

Following the protocol used operationally by the NWCC, the 500-hPa maps assessed in this study are centred over the Northwest geographic area, with a geographical extent of 110°–135°W longitude and 35°–55°N latitude. Geopotential heights are sampled at 30 points across this area at five-degree increments from Global Forecast System (GFS) initialisations at 00Z, which corresponds to 5:00 pm local time (i.e. Pacific Daylight Time) during the fire season. This 30-point grid, which is relatively coarse compared with currently available weather products, was adopted by the NWCC in the late 1980s owing to data availability and computational limitations. This spatiotemporal sampling methodology has been used operationally by the NWCC since the early 1990s and captures much of the synoptic-scale variation in the region.

The 500-hPa geopotential heights for a given day are correlated against a set of 13 operational map type templates (Fig. 3), and the map type that is most highly correlated with a given day’s geopotential heights is selected to represent that day. This set of 13 map types was developed by fire meteorologists at the NWCC to characterise the region’s major synoptic-scale patterns. The map types (UMAPs) can be subdivided into several groups based on their dominant synoptic patterns. UMAP 1 and UMAP 2 both feature offshore upper-level troughs that drive southwesterly winds and precipitation. UMAP 3 through UMAP 6 are characterised by troughs over the Pacific Northwest and cooler temperatures. UMAP 7 and UMAP 8 indicate zonal and split flow, respectively, and suggest onshore winds and mild temperatures. Finally, UMAP 9 through UMAP 13 are ridging patterns, associated with higher temperatures and dry conditions.

Fig. 3.

Thirteen 500-hPa map types used operationally by the NWCC for fire weather forecasting. The 30-point geopotential height data used for matching have been smoothed by taking a composite of 1-degree ERA5 geopotential heights from days that match to each map type over the study period.

Our analyses test 10 standard fire danger indices for both correlation with the 500-hPa map types and their utility in forecasting wildfire occurrence. Four of these, Energy Release Component (ERC), Spread Component (SC), Burning Index (BI) and Ignition Component (IC), are composite fire danger indices that are part of the National Fire Danger Rating System (National Wildfire Coordinating Group 2002). Four are fuel moisture indices, which include 100-hour fuel moisture content (F100) for small-diameter dead fuels, 1000-hour fuel moisture content (F1000) for larger-diameter dead fuels, herbaceous fuel moisture (FMH) for live herbaceous vegetation and woody fuel moisture (FMW) for live woody vegetation. Additionally, we consider daily maximum vapour pressure deficit (VPDM) and the Keetch–Byram Drought Index (KBDI).

Each of the 10 fire danger indices is calculated daily for each PSA. Weather metrics including daily maximum temperature, minimum relative humidity, precipitation duration and VPDM are measured at key RAWSs within each PSA, shown in Fig. 1. There are 72 key RAWSs in the Northwest, with each PSA represented by between three and nine RAWS. The RAWS density is fairly high in most of the Northwest, and the selection of the key RAWSs for each PSA in the Northwest was optimised by NWCC fire weather meteorologists in 2001 (Brown et al. 2011). The RAWS observations serve as inputs to the FireFamilyPlus software (Bradshaw and McCormick 2000), which generates daily fire danger indices for each PSA using equations from the 2016 version of the National Fire Danger Rating System (NFDRS).

To quantify the value added by incorporating map types into predictive models of large wildfire, we establish a baseline using historical 7-Day Significant Fire Potential Outlooks (SFPOs) issued for the Northwest geographic area from 2006 to 2020. Although these outlooks have shifted from a nine-level qualitative scale to a five-level qualitative scale over the history of the SFPO, both coding schemes include a critical threshold at which the expected probability of significant wildfire exceeds 20% on a given day. In our assessment of the value added to fire weather forecasting by incorporating 500-hPa map types, we therefore flatten the qualitative SFPOs to a binary value for each day that indicates whether the forecast probability of a large fire was below or above this 20% threshold.

Statistical approach

To address our first research question, which compares the regional incidence of large fires and widespread fire outbreaks among 500-hPa map types, we consider two outcomes: (i) the probability that a large wildfire will ignite on a given day, and (ii) the probability that a high number of ignitions will occur on a given day. Our analysis of the probability of large wildfire is conducted using PSA-specific size thresholds. That is, if a fire that ultimately exceeds the size threshold for the PSA in which the fire ignites on a given day, that day is designated as a ‘large fire’ day. For the second measure of risk, we define ‘high ignition’ days as those with ignition counts at or above the 80th percentile, which corresponds to days with 33 or more ignitions across the entire GACC. The statistical significance of variation between map types in the frequency of high-ignition days and days with large fires is assessed using a chi-square test.

Our second research question examines the relationship between the 500-hPa map types and fire danger indices. This analysis is conducted at the GACC level by calculating aggregate fire danger for the entire geographic area using the area-weighted mean of daily PSA-level indices. We then estimate Pearson correlation coefficients between a set of dichotomous map type indicators and each fire danger index. These correlations allow us to determine whether there is a statistically significant difference (P < 0.05) in fire danger between days associated with a given map type and days when a different map type is observed.

Our third research question assesses the value of the 500-hPa map types in augmenting subregional models of wildfire risk. In this analysis, we compare the performance of statistical models for two different PSA-level outcomes: (i) the total number of ignitions on each day, and (ii) the probability of a large fire igniting on each day. The models of total daily ignitions are estimated using negative binomial regression, which is appropriate for overdispersed count data and performs well in this empirical context (Arienti et al. 2009). The models of the probability of a large fire igniting on a given day are estimated using logistic regression, which is commonly used to model binary outcomes (Andrews et al. 2003).

For each outcome and PSA, we test a set of 10 fire-danger-only models and a set of 10 with-map-type models. In the fire-danger-only models, the outcome of interest is modelled using each of the 10 fire danger indices, in turn, as a predictor. The 10 with-map-type models build on the fire-danger-only models by pairing each of the 10 fire danger indices with a set of daily map type indicators. After fitting each set of candidate models, we identify the most predictive model by comparing Akaike weights between the models. Akaike weights compare the Akaike information criterion (AIC) of each candidate model against the model with the lowest AIC, estimating the probability that each model provides the best fit within the set of candidate models (Wagenmakers and Farrell 2004).

Finally, we assess the predictive value added by the map types by evaluating how our logistic models of large fire improve on the accuracy of historical SFPOs. We evaluate the forecasts using two metrics: the false alarm rate and the miss rate. The false alarm rate is calculated by dividing the number of false positives (days when large fires were predicted but did not occur) by the total number of days when large fires were predicted. The miss rate is calculated by dividing the number of false negatives (days when large fires were not predicted but did occur) by the total number of days when large fires occurred. We calculate both metrics first for the historical SFPOs and then for forecasts that augment the SFPOs with the output of our logistic models of large fire. The difference between these two sets of metrics represents the operational value added by incorporating map types into forecasts of large wildfire.

Regional association between 500-hPa map types and wildfire risk

At the regional level, there is notable variation between map types in both the probability of observing a high-ignition day and the probability of a large fire igniting, as shown in Fig. 4. The three map types in the top-right quadrant of the graph (UMAPs 9–11) are all ridging patterns associated with elevated probabilities of both high ignition counts and large fires. Statistically significant (P < 0.05) increases in high-ignition days are observed for UMAP 2 and UMAPs 9–11, whereas UMAPs 9, 11 and 12 have significantly more days with large fires. In contrast, the patterns corresponding to upper-level troughs (UMAPs 3–6) have the lowest probability of high ignition counts and large fire. UMAPs 1, 4–6 and 8 have significantly fewer high-ignition days, and UMAPs 4–6 have significantly fewer days with large fires.

Fig. 4.

GACC-level wildfire risk for days when each map type is observed. Vertical axis shows the daily probability of a large wildfire, as defined using operational thresholds, starting when each map type is observed. Horizontal axis shows the probability of a high-ignition day, defined as days with ignition counts at or above the 80th percentile. Numbers within circles indicate daily map type. Size of circles indicates the average number of days that each map type was observed each fire season. The colour of each circle indicates whether the difference between each map type and all other map types is statistically significant (P < 0.05) using a chi-square test. Dashed lines show the probability of large fire and 80th percentile ignition counts on all days.

The map types are also associated with significant variability in regional fire danger indices. Fig. 5 shows the correlations between GACC-level fire danger indices and daily assigned 500-hPa map types. The trough patterns (UMAPs 3–6) are observed to have strongly significant negative associations with the NFDRS fire danger indices (SC, ERC, BI and IC) and generally positive associations with the fuel moisture indices (F100, F1000, FMH and FMW). The ridge patterns (UMAPs 9–13) illustrate the inverse pattern of correlation: these map types are associated with high fire danger and low fuel moisture. The offshore troughs (UMAPs 1–2), zonal flow (UMAP 7) and split flow (UMAP 8) patterns deviate less significantly from average levels of fire danger and fuel moisture.

Fig. 5.

Correlations between map type indicators and GACC-level daily fire danger indices. Map type indicators are coded dichotomously with 1 for days when the map type was observed and 0 for days when a different map type was observed. Note: *** P < 0.01; ** P < 0.05; * P < 0.10.

Augmenting subregional models with 500-hPa map types

Our model selection procedure indicates that the inclusion of map type effects offers a significant improvement in the performance of daily PSA-level ignition count models. The top panel of Table 2 shows the Akaike weights for the fire-danger-only models (rows 1–10) and the with-map-type models (rows 11–20). The best-performing model among the fire-danger-only and the with-map-type models is highlighted in grey for each PSA. For the ignition count models, the preferred fire danger index does not change between the fire-danger-only and the with-map-type models. ERC and VPDM are the preferred fire danger indices for the majority of PSAs, whereas F1000 is most predictive in the NW04 and NW11 PSAs. Finally, the bottom panel of Table 2 shows the Akaike weight comparison between the best fire-danger-only and the best with-map-type model for each PSA. For all PSAs, the models that include map type indicators outperform those that rely on fire danger indices alone.

The model selection results for the probability of large fire are analogously shown in Table 3. VPDM is the most predictive fire danger index for most PSAs in both the fire-danger-only and with-map-type models. As shown in the lower panel of the table, the inclusion of map type improves the prediction of large fire in 4 of the 12 PSAs. In eight of the PSAs, the additional explanatory power of the map type indicators does not outweigh the model complexity penalty imposed by the AIC estimator. In none of the PSAs were statistically significant map type effects observed in models of the probability of large fire.

Map-type effects for the best-performing models of total daily ignitions are estimated as incidence rate ratios and displayed for each PSA in Fig. 6. These incidence rate ratios indicate how many times more (or fewer) ignitions occurred when each map type was observed relative to what would be expected when a different map type occurred on a day with the same level of fire danger. Notably, the PSAs in eastern Washington and all of Oregon have significantly higher expected daily ignitions on days when UMAP 2, the offshore trough, was observed, experiencing between 1.23 and 1.98 times higher ignition counts than expected given the fire danger. The ridging patterns (UMAPs 9–13) were notably heterogeneous in their effects, particularly in southwestern Oregon. Whereas UMAPs 9–11 were associated with an increase in conditional ignitions in the NW04 and NW07 PSAs, UMAP 12 was associated with fewer ignitions than expected in these PSAs.

Fig. 6.

PSA-level incidence rate ratios for the map types, controlling for daily fire danger index, year and day of week. These effects indicate the factor by which observing each map type shifts expected daily ignition counts in each PSA based on fire danger indices alone. The statistical significance of these effects is indicated by hatching (ns, not significant).

Our assessment of the value added by incorporating map types into the forecasting of large fires indicates that our logistic models reduce the miss rate in 8 of the 12 PSAs, while increasing the false alarm rate in only five PSAs. Fig. 7 shows the false alarm and miss rates for both the historical SFPOs and a forecast constructed by taking the union of the SFPO and with-map-type logistic models of large wildfire. That is, a large fire is forecast if either the SFPO or our logistic models incorporating map types indicate a 20% or greater predicted probability of large fire. The horizontal lines in Fig. 7 show the direction of change in the false alarm and miss rates from the historical SFPOs to the forecasts constructed by combining the SFPOs with the output of our logistic models. Reductions in the false alarm and miss rates are plotted in green and represent positive values added by the logistic models, whereas increases in the false alarm and miss rates are shown in purple. The greatest reduction in the miss rate is achieved in NW11, where the logistic models correctly identify six large wildfires that the SFPOs missed out of the 66 total.

Fig. 7.

Comparison of false alarm and miss rates for large fires between historical Significant Fire Potential Outlooks (SFPOs) and predictions that also incorporate the results of the with-map-type logistic models presented here, for each of the 12 predictive service areas in the Northwest (NW01–NW12). Reductions in the false alarm and miss rates represent positive value added by the logistic models and are indicated by green lines. Increases in the false alarm and miss rates due to the incorporation of the logistic models are shown in purple.