Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2021 Sep 2.
Published in final edited form as: Int J Wildland Fire. 2014 Nov 18;23(8):1119–1129. doi: 10.1071/wf13046

Santa Ana winds and predictors of wildfire progression in southern California

Michael Billmire A,D, Nancy H F French A, Tatiana Loboda B, R Chris Owen A,C, Marlene Tyner A
PMCID: PMC8411922  NIHMSID: NIHMS1726022  PMID: 34483633

Abstract

Santa Ana winds have been implicated as a major driver of large wildfires in southern California. While numerous anecdotal reports exist, there is little quantitative analysis in peer-reviewed literature on how this weather phenomenon influences fire progression rates. We analysed fire progression within 158 fire events in southern California as a function of meteorologically defined Santa Ana conditions between 2001 and 2009. Our results show quantitatively that burned area per day is 3.5–4.5 times larger on Santa Ana days than on non-Santa Ana days. Santa Ana definition parameters (relative humidity, wind speed) along with other predictor variables (air temperature, fuel temperature, 10-h fuel moisture, population density, slope, fuel loading, previous-day burn perimeter) were tested individually and in combination for correlation with subsets of daily burned area. Relative humidity had the most consistently strong correlation with burned area per day. Gust and peak wind speed had a strong positive correlation with burned area per day particularly within subsets of burned area representing only the first day of a fire, >500 ha burned areas, and on Santa Ana days. The suite of variables comprising the best-fit generalised linear model for predicting burned area (R2 = 0.41) included relative humidity, peak wind speed, previous-day burn perimeter and two binary indicators for first and last day of a fire event.

Additional keywords: chaparral, fire spread, generalised linear model, relative humidity, wind speed

Introduction

The devastating 2003 and 2007 southern California wildfires were reportedly driven by a combination of extremely flammable chaparral fuels and Santa Ana (SA) winds (Westerling et al. 2004; Miller and Schlegel 2006; Keeley et al. 2009). The relationship between these strong, dry winds and high-severity wildfire in the dense and continuous chaparral fuels of southern California is intuitive and anecdotally supported. However, there is a paucity of rigorous quantitative analyses in the peer-reviewed literature to support the relationship. The goal of this study is to provide a quantitative assessment of the effect of meteorologically defined SA events on wildfire progression rates in southern California. The results of this assessment can be directly applied to development of policies related to fire-fighter and public safety as well as supporting fire spread modelling and fire ecology studies.

Santa Ana events

SA events are a seasonal weather phenomenon affecting southern California from Los Angeles to San Diego (Schroeder et al. 1964). SA events typically last 1.5 days (Raphael 2003) and occur between September and April, with various indices of SA event frequency showing that December and November experience the highest number of SA events (Raphael 2003; Hughes and Hall 2010). The strong winds associated with SA events are driven by a combination of pressure and temperature gradient anomalies between the Great Basin and the ocean off the coast of southern California (Raphael 2003; Hughes and Hall 2010).

A complicating factor in the study of SA events is the lack of standardised criteria for the defining meteorological parameters and their values. In the published literature, SA events have been defined in terms of various combinations and thresholds of barometric pressure, temperature, humidity, and wind speed and direction. For example, Edinger et al. (1964) based their definition of a SA event on ‘the occurrence of a 3-mb drop in pressure from Palmdale to Santa Monica, California; a 9°C or greater increase in temperature from Santa Monica to Palmdale; northerly winds with speeds of 13 m s−1 or greater in the Riverside, San Bernadino Valley, California; and a relative humidity of 30% or less in the Los Angeles, California area’. Sergius and Huntoon (1956) characterised SA events as having winds coming from the north-east and averaging 9 m s−1 or greater on at least four hourly observations while the relative humidity was less than 40%. Richardson (1973) characterised SA events only by the pressure gradient between Las Vegas, Nevada and San Diego, CA. Moritz (1997) used a temperature threshold of 32°C at the Santa Barbara airport (based on Davis and Michaelsen 1995). Hughes and Hall (2010) used cluster analysis to identify SA winds and classify them into intensity categories for wind speeds >5 m s−1. San Diego County meteorologists have identified SA winds as those with a wind direction of 345–115°, wind speed >3.6 m s−1 and relative humidity <30% (W.C. Brick, San Diego Air Pollution Control District, pers. comm.). Because of this lack of consensus, we evaluate several different SA definitions in the present study.

Santa Ana events and wildfire

A growing body of research has established a largely circumstantial link between SA events and wildfire occurrence and size. It has been shown that most of the largest and longest duration wildfires in southern California have historically occurred at the start of the SA season (Davis and Michaelsen 1995; Keeley and Fotheringham 2001; Westerling et al. 2004), and further that they are often accompanied by SA events (Keeley and Zedler 2009; Keeley et al. 2009). However, quantitative links are sparse. A weak positive correlation between annual total burned area and number of SA days in southern California has been demonstrated via model simulation (Peterson et al. 2011) and empirical data (Price et al. 2012). The frequency of large (>1000 ha) fires has also been tied to SA events (Davis and Michaelsen 1995; Moritz 1997), with the most solid quantitative link established by Moritz et al. (2010), who showed that from 1995 to 2003, >1000-ha wildfires were more likely to occur in areas of high fire risk according to the Fosberg Fire Weather Index (FFWI) (Fosberg 1978) averaged across October SA event days. There remains a distinct need to evaluate the effect of SA events on wildfire size and progression at finer temporal and spatial scales.

Predictors of wildland fire progression

Despite the paucity of quantitative analysis on SA events and fire, empirical studies modelling the influence of SA event components (relative humidity, wind speed, wind direction) and other predictors of fire rate of spread (ROS) are abundant. Indeed, the body of literature concerning ROS modelling is extensive, and is excellently summarised by Sullivan (2009). Among empirical models, wind speed and fuel moisture have been shown most often to be the predominant drivers of ROS (Sullivan 2009). Other parameters that influence ROS include relative humidity (Lindenmuth and Davis 1973; Fernandes 2001; Zhou et al. 2005), wind direction in relation to fire front orientation (Cheney and Gould 1995), fuel loading (Burrows et al. 1991; Bilgili and Saglam 2003), air temperature (Lindenmuth and Davis 1973; Fernandes 2001; Weise et al. 2005; Zhou et al. 2005), fuel temperature (Lindenmuth and Davis 1973), topographic slope (Rothermel 1972; Fernandes et al. 2002; Linn et al. 2010) and ignition length (Cheney and Gould 1995). Due to the inherent difficulty of measuring ROS of large wildfires in situ these empirical models are based primarily on observations of low-intensity prescribed burns on <1-ha plots with occasional validation or data fitting using well-documented wildfires (Sullivan 2009). As prescribed burns cannot fully reproduce the dynamics of wildfires (Fendell and Wolff 2001; Coen and Douglas 2010), particularly those of the size and severity that periodically occur in southern California, there is a continued need for data on the predictors of in situ wildfire ROS against which to compare these models.

One difficulty in conducting quantitative analysis on observed wildfire ROS to this point has been the impracticality of measuring ROS in a wildfire setting. A solution that has been used in limited fashion for well-documented fires, such as those that occurred in Yellowstone National Park in 1988 (Turner et al. 1994), is to measure two-dimensional burned area increase over iterated time intervals with remote sensing or field-based mapping methods. Burn perimeter data are prevalent and publicly available, but are often either spatially non-comprehensive (i.e. based on field notes or limited satellite overpass data; e.g. Moderate Resolution Imaging Spectroradiometer (MODIS) burn products) or temporally variable (i.e. representing a burn event lasting anywhere from 1 day to several weeks; e.g. Monitoring Trends in Burn Severity (MTBS) Eidenshink et al. 2007). For this study, we combine MODIS fire products with MTBS burn perimeters to create spatially complete and temporally consistent wildfire progression estimates that allow for a spatially explicit assessment at a daily temporal resolution. As ROS is usually measured at a finer spatial and temporal scale (in units of m s−1) but our study concerns the measurement of two spatial dimensions and a lower temporal resolution (in units of ha day−1), we refer to our data as reflecting fire progression rather than ROS.

Objectives

The vulnerability of southern California to extreme SA-influenced wildfire events makes it imperative that we develop a more detailed and quantitative understanding of the role of SA events and other predictors of wildfire progression. Our first objective in this study was to conduct a rigorous quantitative evaluation of the direct effect of SA events on the amount of area burned in a day during a wildfire event. Our second objective was to evaluate the individual components that define SA events and characterise other predictors of daily wildfire progression rates in southern California.

Methods

Fire progression dataset

For this study, we utilised a fire progression algorithm that enabled an assessment of wildfire progression rates at a daily time scale. This algorithm was developed as a semi-automated approach suitable for developing daily estimates of area burned from satellite observations of fire occurrence. Fire progression was based on observed active fire detections from both the Terra and Aqua satellites using the MODIS active fire product (Giglio et al. 2003). Fire progression was assessed at a daily time step and was based on the first observation of fire occurrence in a particular MODIS 1-km pixel within a burn scar. To define known burned scars and develop the most comprehensive database of burned areas within the study region, we combined burn perimeters from the MTBS database with the MODIS direct broadcast burned area maps (MODIS MCD64A1, Giglio et al. 2009). The MTBS burn perimeters provided the basis for the combined product, which was enhanced by including scars from the MODIS direct broadcast burned area product that were unmapped in the MTBS dataset. We focused the algorithm on modelling the approximate date of burning within the combined burned area product using the date–time information contained in the MODIS active fire product and adjusted for the local standard time correction.

Active fire detections were first processed using the Fire Spread Reconstruction (FSR) approach, which clusters individual fire points in space–time to identify contiguous fire events and groups of fire events creating a single burn scar (Loboda and Csiszar 2007). The clustered fire events were subsequently processed to separate the first date of fire detection within a specific MODIS pixel from subsequent detections, wherein the pixel continues to emit a sufficient amount of thermal energy to warrant the detection of ongoing burning. The first group is further compared against the known burned areas, and if the number of those fire detections multiplied by 1 km2 (which represents the nominal resolution of the MODIS pixel in thermal bands) provides an area estimate ≥80% of the actual mapped scar, they were then used to interpolate fire progression surfaces by date of burning. The date of burning between adjacent dates of fire detections in fire progression surfaces was interpolated using the inverse distance mapping algorithm with a radius of 3 km. In cases where burn scars did not overlap with the MODIS active fire detections, date of burning was assigned from the start date for MTBS database burns and from the date of mapping for the MODIS burned areas.

Figure 1 shows an example fire event (the 2003 Cedar Fire) with daily progression as determined by the algorithm described above. Reliable validation data are limited, but the daily progression compares well with real-time burned area estimates recorded by fire incident managers (ICS-209, NWCG 2011) for the example shown (2003 Cedar Fire, Fig. 2). We note that the comparison is complicated by uncertainties in ICS-209 area estimates due to both typographical errors (Raffuse et al. 2008) and variability in methods used by the managers to estimate area. Additionally, the ICS-209 reporting schedule does not typically align with the midnight daily cutoff used in our methods. Wildfire progression mapped with the FSR algorithm was previously found to be consistent with field measurements in boreal forests (Loboda and Csiszar 2007).

Fig. 1.

Fig. 1.

Open in a new tab

Example burned area progression polygons comprising a single fire event calculated using Fire Spread Reconstruction-based methods described in Loboda and Csiszar (2007) with modifications specific to this study. The example shows the Cedar Fire that occurred in October 2003 San Diego County, CA, USA.

Fig. 2.

Fig. 2.

Open in a new tab

Comparison of daily burned area estimates for the October 2003 Cedar Fire that occurred in San Diego County, California, USA. ICS-209 = Incident Status Summary Form. FSR = Fire Spread Reconstruction; methods described in Loboda and Csiszar (2007) with modifications specific to this study.

The fire progression dataset used in this study contained 528 burned area polygons grouped by fire event and date. The data comprised a set of 158 distinct fire events (67 single-day and 89 multi-day fires) that were detected by fire management agencies and MODIS active fire products, had a size >1 km2, and occurred from 2001 to 2009 in southern California within32.5 to 35.4° latitude and −116.0 to −120.6° longitude. Where indicated by MTBS data or ICS-209 reports (NWCG 2011), we removed prescribed burns so that the relationships we tested did not reflect aggressively controlled fire behaviour. Burned area per day was non-parametrically distributed with a mean of 2081 ± 5055 ha (standard deviation) and a median of 635 ha. Individual fire events in the dataset lasted for a mean of 3.4 days for all fires and a mean of 5.5 days for multi-day fires alone. According to Fuel Characteristic Classification System (FCCS) maps (USGS 2010) and not counting areas that burned on multiple instances within the period, the collective burned area comprised 72% chaparral or other shrublands, 17% forest or woodland and 6% grassland, with the remaining 5% distributed among agriculture, urban and riparian cover types.

Santa Ana event criteria

The three a priori definitions of SA events used to conduct this analysis were selected because they are classified using wind direction, wind speed, and relative humidity (Table 1), all of which are reported by Remote Automated Weather Stations (RAWS) (National Interagency Fire Center, see http://raws.fam.nwcg.gov, verified 23 September 2014). The first definition, SA1, is a classification used by San Diego County meteorologists to identify SA wind events, and combines wind direction, speed and relative humidity thresholds. Definition SA2 is the National Weather Service’s Red Flag Day criterion for the San Diego Weather Forecast Office and is based on sustained very high wind gusts and very low relative humidity levels. It has the most extreme thresholds of the three definitions, but is not specific to SA wind events as there is no wind direction component. The third definition, SA3, was defined by Sergius and Huntoon (1956) and is one of the more commonly used SA event criteria in the peer-reviewed literature. It combines wind direction, wind speed and a late afternoon relative humidity threshold. SA1 is the least restrictive of the three definitions; over the study period, all SA2 and SA3 detections were also detected via SA1 criteria.

Table 1.

Santa Ana event classifications based on wind direction, wind speed and relative humidity

ID Source Wind direction Wind speed Relative humidity
SA1 San Diego CountyA >345° and <115° true >3.6 m s−1 <30%
SA2 National Weather Service Red Flag Day Criteria for San Diego Weather Forecast Office – >11.2 m s−1 sustained for ≥6 h <15% for >6 h
>15.6 m s−1 frequent gusts for ≥6 h
SA3 Sergius and Huntoon (1956) 0–90° true ≥8.9 m s−1 on ≥4 separate hourly recordings <40% at 1630 hours PST
Open in a new tab
A

W. C. Brick, San Diego Air Pollution Control District, pers. comm.

Remote Automated Weather Station data

We elected to use station-based weather data versus a gridded dataset because RAWS units are well distributed throughout our study area (Fig. 3), and the station-based data retained the actual hourly measured values for weather variables instead of substituting them for modelled (either through direct spatial interpolation or other statistical techniques) values. Hourly measurements are required for the SA2 and SA3 definitions.

Fig. 3.

Fig. 3.

Open in a new tab

Overview of the southern California study area showing locations of burn progression polygons and Remote Automated Weather Stations (National Interagency Fire Center, see http://raws.fam.nwcg.gov).

We compiled and processed hourly data from 82 RAWS units (made available through MesoWest) in southern California from 2001 to 2009, from which we derived data for 15 variables; 10 taken directly from RAWS and five secondary variables computed from RAWS data. Table 2 catalogues variable abbreviations and descriptions. We only used data that were flagged as ‘OK’ by MesoWest’s quality control process (MesoWest Quality Control Flags Help Page, see http://mesowest.utah.edu/html/help/key.html, verified 23 September 2014). From RAWS we directly extracted measurements of relative humidity (RH), wind speed (WS), wind direction, fuel temperature, air temperature and 10-h fuel moisture (FM). Note that many RAWS stations do not collect data for all variables at all times. Of the 528 burned areas, 501 had gust WS data, 270 had peak WS data, 509 had fuel temperature data and 207 had FM data. All WS measurements for RAWS units are taken at 20 ft (6.1 m). Variables derived from RAWS data include the FFWI and four SA classification indicators: true or false flags for SA1, SA2 and SA3, and daily total number of hours that met SA1 criteria (see Table 1 for SA definitions).

Table 2.

Parameter abbreviations and definitions for the independent variables assessed in this analysis

Parameter Description
RHA Mean relative humidity (RH) (%)
RH (daily minimum)A Daily minimum RH (%)
RH (at 1600 PST)A Used to approximate the RH at 1630 hours PST criteria for SA3
WSA Mean 20-ft (6.1 m) wind speed
WS (gust)A Mean gust speed (defined as the maximum 3-s mean over a 2-min period)
WS (peak)A Mean peak wind speed
FTA Mean fuel temperature
ATA Maximum air temperature
FMA Mean 10-h dead fuel moisture
SA1 h (#) Number of hours (0–24) in a day meeting SA1 criteria
PD Mean US CENSUS 2000-derived population density
HD Mean US CENSUS 2000-derived housing unit density
LP The length of the burn perimeter from the previous day
SL Topographic slope, derived from the National Elevation Dataset 1/3 arcsecond (10-m) data
FL Mass per area summed across all fuel strata from the 30-m Fuel Characteristic Classification System map (Ottmar et al. 2007)
FFWI Fosberg Fire Weather Index (Fosberg 1978), based on WS, air temperature and RH
Open in a new tab
A

Data derived from the nearest (by straight line) quality-checked Remote Automated Weather Station (National Interagency Fire Center, see http://raws.fam.nwcg.gov). ‘Mean’ values are daily mean of hourly measurements.

We derived weather data for each burned area by linking each burned area to the nearest RAWS unit by straight-line distance. In order to maximise our sample size, if the data from RAWS unit nearest to a particular burned area polygon was missing data or did not meet MesoWest’s quality standards for that polygon’s burn date, we assigned the second nearest RAWS unit, and likewise we used the third nearest unit if the second nearest did not meet quality standards. The mean distance from a burned area polygon to its assigned RAWS unit was 8.5 ± 6.4 km.

Since SA events are regional in scale, they were flagged for the entire study region on a daily basis rather than according to local conditions for each burned area. To reduce false positive detections caused by transient local conditions at one or two individual RAWS units, days were flagged as SA if greater than 5% of reporting RAWS units detected SA conditions. We use a percentage rather than an absolute number of RAWS units because the number of RAWS units with reported data varied daily from 37 to 74 across the study time period. A 5% threshold guaranteed that at least two RAWS units detected an SA event on the low-data days and at least four detected an SA event on higher-data days. See Results for additional justification of the 5% threshold.

Study variables

Burned area per day was the dependent variable evaluated in all analyses. In order to tease out more nuanced relationships with independent variables, we also evaluated subsets of burned area per day based on duration, temporal position within the fire event, burned area size class and SA criteria (Table 3). For temporal position within the fire event, we separately evaluated burned areas representing only the first day of a multi-day fire event (first-day-of-fire-event; e.g. for a fire lasting 4 days from 1–4 October, the first day would be 1 October), middle day (middle-day-of-fire-event; 2–3 October) and last day (last-day-of-fire-event; 4 October). We evaluated these aggregate subsets as opposed to incremental daily (i.e. day1, day2,…, daylast) subsets because the large range in fire event duration (1 day to several weeks) would have left most subsets with insufficient sample sizes.

Table 3. Daily burned area (ha) across 158 distinct wildfire events in southern California from 2001 to 2009.

The 158 district wildfire events include wildfire events that were recorded by the MODIS MCD64A1 (Giglio et al. 2009) and Monitoring Trends in Burn Severity datasets (Eidenshink et al. 2007) and exceeded 1 km2. Daily burned area was estimated using the Fire Spread Reconstruction algorithm (Loboda and Csiszar 2007). nSA, the number of daily burned areas in the subset on which SA events (by any definition) occurred

Subset nSA n Mean (ha) STD
All 100 528 2081 5095
Fire event duration
 Single-day events 16 67 948 813
 Multi-day events
 All multi-day 84 461 2250 5435
 First day 27 89 2934 4456
 Middle day 41 283 2528 6363
 Last day 16 89 677 1165
Burn area size
 <500 ha 25 240 140 150
 ≥500 ha 75 288 3706 6472
Santa Ana eventsA
 SA1 – 100 4874 9693
 SA2 – 25 6485 11131
 SA3 – 70 6388 11180
 Any – 100 4874 9693
 None – 428 1433 2809
Open in a new tab
A

See Table 1 for Santa Ana event definitions.

We evaluated a suite of variables that were expected to influence fire progression according to implications from previous studies (see Introduction). We included any and all measurable variables that might influence fire progression in order to maximise the amount of variability we could explain using multivariate analysis. In addition to the RAWS-derived weather variables, these included previous-day-burn-perimeter (LP), population density (PD), housing unit density (HD), topographic slope and fuel loading.

To account for intuitive and previously quantified effects of ignition length on ROS (Cheney and Gould 1995), we included a parameter to approximate an ignition length metric. For each fire event, we expected that the amount of active burning at the start of a day will influence the total burned area. Thus, for multi-day fire events, we used LP as an approximation for the amount of actively burning fire at the start of the day. For areas representing the first day of a fire, including all single-day fires, this metric was set to zero.

We calculated the area-weighted mean PD and HD based on US Census 2000 tracts and data (US CENSUS Bureau 2002); area-weighted mean slope (degrees) from 1/3 arcsecond (10-m resolution) National Elevation Dataset (USGS 1999); and fuel loadings from the 30-m FCCS map developed by the USDA Forest Service Fire and Environmental Research Applications (FERA) team (Ottmar et al. 2007). Fuel loadings represent mass per area summed across all fuel strata.

Statistical analyses

All datasets and their transformations were tested for parametric distribution using the one-sample Kolmogorov–Smirnov test. Parametric distribution was preferred so that we could utilise more powerful parametric statistical tests (e.g. Pearson’s instead of Spearman’s correlation coefficient). To approximate parametric distributions, we power transformed each non-parametrically distributed dataset by 0.2.

We used Student’s t-tests to determine whether mean burned area per day under SA conditions was greater than that under non-SA conditions. Pearson’s correlation was used to evaluate linear relationships of predictor variables to burned area per day. We used generalised linear model (GLM) analysis to determine the suite of continuous and binary variables that could best predict mean burned area per day. For the binary variables, data consisted of 0 (false) and 1 (true). The binary variables tested included: wind-direction-is-NE (0–90°); is-first-day-of-fire-event; is-last-day-of-fire-event; is-part-of-a-multiday-burn; and is-coastal (within 1 km of the coast). The latter was included based on an outlier analysis that identified two >2000 ha burned areas that occurred on days of particularly high RH (>80%; corroborated by ICS-209 reports) that were both located on the Pacific coast.

Variable selection for the GLM was via all subsets regression (for up to five variables) and best judgment with guidance from the univariate analysis results. Only one RH (daily mean) variable, one WS (peak) variable, one temperature (air temperature) variable and one PD (population) variable were used in the variable selection process, in order to avoid multicollinearity. Standardised coefficients provide an approximate measure of the relative importance of the predictor variables and were calculated for the best-fit GLM by scaling the continuous input variables by two standard deviations. This method follows recommendations by Gelman (2008) for models that include binary predictor variables.

Results

Santa Ana event classifications

The average number of days per month on which SA events were detected by at least 5% of RAWS units in the study area shows a comparable distribution to that of Raphael (2003), who identified SA events from 1968 to 2000 using barometric pressure gradients from daily weather maps (Fig. 4). SA3, the median between the most (SA2) and least (SA1) strict classifications, shows a very similar magnitude to that of Raphael (2003), of around four SA events per month in the peak season of November and December, whereas the less-strict SA1 identifies 6–7 events per month during the same period. Increasing the percentage of RAWS unit detections required to flag an SA event from 5 to 15% produced a time series wherein of the three classifications SA1 became most closely tuned to Raphael (2003), whereas SA3 detections were roughly halved. The strong seasonality trend shown by the three classifications is consistent with that of SA events identified using a variety of alternate criteria (Raphael 2003; Conil and Hall 2006; Hughes and Hall 2010).

Fig. 4.

Fig. 4.

Open in a new tab

Monthly frequency of Santa Ana events in southern California according to various Santa Ana definition criteria: SA1, SA2, and SA3 are defined in Table 1 and were classified using 2000–2009 Remote Automated Weather Station (National Interagency Fire Center, see http://raws.fam.nwcg.gov, verified 23 September 2014) data. Raphael (2003) data are based on Santa Ana events identified using barometric maps from 1968 to 2000. Note that Raphael (2003) data represent the number of Santa Ana events per month (with events lasting 1.5 days on average) as opposed to the number of days on which Santa Ana events occurred.

Burned area during Santa Ana events

Mean burned area per day was 3.4, 3.5 and 4.5 times greater (P < 0.001) on days meeting SA1, SA2 and SA3 criteria (Fig. 5, white boxes) than on days that did not (Fig. 5, grey boxes). Analogous relationships between SA day and non-SA day mean burned areas were found when looking at only single-day fires, only multi-day fires or all other subsets listed in Table 3 (results not shown). By increasing the percentage of RAWS unit detections required to flag an SA event from 5 to 15% (thereby tuning monthly SA1 distribution to that produced by Raphael 2003), mean burned area was found to be 4.0 times greater (P < 0.001) on SA1 days than on non-SA1 days and 3.0 times greater (P < 0.001) on SA3 days than on non-SA3 days. There were no days on which 15% of RAWS met SA2 criteria.

Fig. 5.

Fig. 5.

Open in a new tab

Statistical comparison (t-test) of mean daily burned area on Santa Ana (‘True’) v. non-Santa Ana (‘False’) days. The data represent southern California, USA, for the years 2001–2009. The black bar within the boxes shows the median. The whiskers extend to 1.5 multiplied by the interquartile range of the box if data points exist that far; otherwise they extend to the most extreme data points. The circles represent outliers that do not fit into the whisker range as defined above. Definitions SA1, SA2 and SA3 refer to three different criteria for classifying Santa Ana events as defined in Table 1.

Individual parameter correlations with burned area per day

Univariate correlation analysis results are shown in Table 4. Among all variables that were evaluated, mean RH was the most consistently strong predictor of burned area per day across all subsets of burned area. At least one of either gust or peak WS were positively correlated with burned area per day for most subsets except for last-day-of-burn-event, <500 ha burns and non-SA burns. The first-day-of-fire-event subset had a much stronger correlation strength with WS (r = 0.42, P ≤ 0.001 for gust WS; r = 0.61, P ≤ 0.001 for peak WS) than other subsets.

Table 4. Pearson correlation coefficients (r) evaluating 16 parameters with subsets of wildfire burned area per day in southern California, USA, from 2001 to 2009.

All values listed are significant at P < 0.05; ns = not significant (P ≥ 0.05). The daily burned area dataset comprises 158 distinct wildfire events that were detected by the MODIS MCD64A1 product (Giglio et al. 2009) and Monitoring Trends in Burn Severity datasets (Eidenshink et al. 2007) and exceeded 1 km2.

Parameter All Single-day fire events Multi-day fire events Burned area size Santa AnaB
All multi-day First day Middle day Last day <500 ha >500 ha non-SA SA
n 528 67 461 89 283 89 240 288 428 100
RH −0.47** −0.36 −0.48** −0.46** −0.56** −0.24 −0.31** −0.31** −0.43** −0.26*
RH (daily minimum) −0.43** −0.25 −0.45** −0.42** −0.51** −0.27 −0.32** −0.23** −0.41** ns
RH (at 1600 hours PST) −0.43** −0.32* −0.44** −0.44** −0.49** −0.30* −0.33** −0.22** −0.38** −0.25
WS ns ns ns 0.28 ns ns ns ns ns ns
WS (gust) 0.17** 0.26 0.18** 0.42** ns ns ns 0.18* ns 0.21
WS (peak) 0.25** 0.37 0.25** 0.61** 0.17 ns ns 0.17 ns ns
FT 0.18** ns 0.20** ns 0.28** ns 0.40** ns 0.27** ns
AT 0.18** ns 0.20** ns 0.31** ns 0.40** ns 0.29** ns
FM −0.36** ns −0.38** −0.29 −0.46** ns −0.39** ns −0.42** ns
SA1 h (#) 0.27** 0.36* 0.27** 0.48** 0.24** ns ns 0.22** ns 0.20
PD ns ns ns 0.32** ns 0.35** ns ns ns 0.30*
HD ns ns ns 0.35** ns 0.40** ns ns ns 0.28*
LP 0.18** ns 0.23** ns 0.65** 0.44** ns 0.16* 0.29** ns
SL 0.16** ns 0.18** 0.21 0.18** ns ns ns 0.21** ns
FL 0.24 ns 0.26** ns 0.31** 0.21 0.15 0.13 0.29** ns
FFWI 0.17** 0.28 0.17** 0.34** ns 0.21 ns 0.12 ns ns
Open in a new tab
B

SA refers to daily burned areas meeting any of the three definitions detailed in Table 1.

*

P ≤ 0.01;

**

P ≤ 0.001.

See Table 2 for variable names, descriptions and sources

FM had a negative correlation with burned area per day across all subsets, although correlation strength was generally lower than that of RH. Fuel temperature, air temperature, topographic slope, fuel loadings and FFWI all had very weak positive correlations. The strongest correlation (r = 0.65, P ≤ 0.001) was with LP for the middle-day subset of burned areas. Burned area on non-SA days is correlated with RH, air and fuel temperature, FM, LP, slope and fuel loading, whereas burned area on SA days was correlated with only RH, gust WS, FM, PD and HD.

The dominant predictors of fire progression among first-day-of-fire-event burned areas differed from those for last-day-of-fire-event burned areas. For the former, burned area variability was best predicted by WS and RH. Middle-day burned area was best predicted by RH and LP and last-day burned area was best predicted by LP, and PD and HD. Although fire size category did not affect correlation strength of RH and fuel loading, smaller (<500 ha) fires were predicted by air and fuel temperature and FM whereas larger (>500 ha) fires were predicted instead by gust and peak WS, LP and FFWI.

The best-fit combination of variables for the entire dataset is shown in Table 5. The combination of RH, peak WS, LP, is-first-day-of-fire-event and is-last-day-of-fire-event comprise the best-fit suite of variables, explaining 41.9% deviance in burned area per day. Standardised coefficients show that LP and is-first-day-of-fire-event were the most influential predictors followed by RH, is-last-day-of-fire-event and finally peak WS.

Table 5. Best-fit generalised linear model for predicting burned area per day for wildfires in southern California, USA, 2001–2009.

Dependent variable units are square metres per day power transformed by 0.2. Continuous independent variables (relative humidity, peak wind speed, previous-day-burn-perimeter) were also power transformed by 0.2. Variable selection was conducted using all subsets regression (for up to five variables) and best judgment based on separate univariate analysis results. Standardised coefficients provide an approximate measure of the relative importance of the predictor variables and were calculated by scaling the continuous variables (relative humidity, peak wind speed, previous-day-burn-perimeter) by two standard deviations following recommendations by Gelman (2008) for models that include binary predictor variables (is-first-day-of-fire-event, is-last-day-of-fire-event). Variables tested as possible predictors that were not components of the best-fit model included air temperature; FCCS fuel loading; population density (2000 USA Census); topographic slope; fuel temperature; burn-area-is-within-1-km-of-coast; wind-direction-is-from-the-north-east and burn-is-part-of-a-multiday-fire

Parameter Coefficient Standardised coefficient P-value
Relative humidity (daily mean %) −9.69 −4.8 <0.001
Peak wind speed (mean of hourly, m s−1) 4.28 1.8 0.068
Previous-day-burn-perimeter (m) 1.55 13.3 <0.001
is-first-day-of-fire-eventA 12.27 12.3 <0.001
is-last-day-of-fire-eventA −4.82 −4.8 <0.001
Intercept 25.24 <0.001
R 2 0.408
Deviance explained 41.9%
Open in a new tab
A

For example, for a fire event that lasted from 1–4 October, the first day is 1 October and the last is 4 October.

Discussion

Effects of Santa Ana events on wildfire

The results showing that wildfires occurring on SA days are 3.5–4.5 times larger than those occurring on non-SA days decisively confirm the intuitive causal relationship between SA events and large wildfires. These findings, along with studies showing correlation between annual burned area and number of SA days (Peterson et al. 2011; Price et al. 2012) and linking FFWI of SA events to large wildfires (Moritz et al. 2010), help to further quantitatively characterise the effects of SA events on wildfire size. Although we did not find a significant correlation between FFWI and burned area on SA days, as would be expected based on Moritz et al. (2010), there was a weak though far from statistically significant correlation (r = 0.11, P = 0.27). Further, the significant positive correlation with FFWI among the large area (>500 ha) subset of burned areas (Table 4) is consistent with the results of that study.

Univariate results (Table 4) show that the set of predictors of fire progression on non-SA days consists of RH, fuel, LP, and air and fuel temperature whereas SA wildfires are predicted by RH, WS, and PD and HD. The relative weakness in correlation strength between the subset of burned area on SA days and RH and gust WS is primarily a consequence of using thresholds based on those variables to define the subset; effectively, the observed data range is greatly reduced, and this makes it more difficult to discern strong statistical relationships. The PD and HD relationships are almost certainly an effect of SA events rather than a driver; SA winds push wildfires from inland towards the denser population centres on the coast. The larger a burned area becomes, the more likely it is to reach the more densely populated areas. Wind direction trends support this explanation: wind direction on SA days averaged 45° while on non-SA days winds originated from the south-west (averaging 235°), blowing towards less densely populated areas.

Santa Ana event definitions

The National Weather Service Red Flag Day criterion for the San Diego Weather Forecast Office (i.e. SA2) previously consisted of only a low RH threshold, but in spring 2010 the criteria were redefined by adding high WS thresholds in order to reduce the frequency of warnings (The San Diego Union–Tribune, see http://www.utsandiego.com/news/2010/apr/15/officials-hope-policy-shift-will-save-money-and-li/, verified 23 September 2014). This change was reportedly made because the warnings were so common that they had started to become ‘white noise’ to the public. On the surface, our results suggest that the high WS thresholds may be slightly too strict: the largest burned area in our dataset (26 October 2003 of the Cedar fire) was not captured by SA2 (Fig. 5), and the less-strict SA3 definition captured a larger effect (4.5 times greater than non-SA3) on burned area per day than SA2 (3.5 times greater than non-SA2). However, a closer examination showed that 26 October 2003 very likely would have been flagged by SA2 criteria except that many RAWS units did not report data for several hours in the middle of the day when winds reached their peak. On that day, RAWS units reported an average of only 17.6 hourly measurements (compared to 22.7 hourly measurements per day across the entire study time period). Had 26 October 2003 been flagged as SA2, the Cedar burn and 10 additional burned areas occurring that day would have counted as SA2, and mean burned area per day under SA2 would have increased to 5.1 times that of non-SA2 burned areas, exceeding the effect of the SA3 events. However, less-strict parameter thresholds such as those of SA3 still capture a very large effect on burned area. Perhaps a middle ground exists to balance between the potential effect on wildfire and avoiding ‘crying wolf’: at least 5% RAWS units reported SA3 conditions on 25 days per year whereas SA2 conditions were reported on only 5.5 days per year on average from 2000 to 2009.

Although not necessarily having the same public safety implications, the absence of standardised values for SA event meteorological conditions remains somewhat of an obstacle to studies of SA events. Raphael (2003) arguably provides the most direct method of identifying SA events by using barometric maps to detect the characteristic pressure gradients that in turn cause the humidity and wind conditions characterised in this study. This method, however, is difficult to automate because barometric data are less commonly available in distributed weather datasets than RH, WS and wind direction data, and the detection of the characteristic pressure gradient requires either manual delineation or more advanced spatial analyses. Thus, it would be advantageous to examine which definition (SA1, SA2, SA3 or others) and RAWS unit detection threshold (i.e. ≥5%) best align with daily SA event detections using barometric maps. Our results comparing only aggregate monthly SA detections to those published by Raphael (2003) (Fig. 4) suggest that SA events may be best characterised either by a few detections of higher-threshold criteria (i.e. 5% of RAWS units detecting SA3) or a more regional detection of lower-threshold criteria (i.e. 15% of RAWS units detecting SA1).

Predictors of wildfire progression

The univariate and GLM results help to tease out the components of SA events as well as additional variables that contribute most to wildfire progression. The 41% explained deviance for our best-fit GLM (Table 5) is low compared with laboratory and prescribed burn studies of shrubland fuels (McCaw 1997; Catchpole et al. 1998; Baeza et al. 2002; Bilgili and Saglam 2003), but it is not entirely unexpected given the complexity of in situ wildfire dynamics. That the model consists of a WS measurement and a moisture indicator (RH) is consistent with the characteristics of SA events as well as previous models based on prescribed burns (McCaw 1997; Fernandes 2001). The inclusion of a variable approximating ignition length (previous-day burn perimeter) is consistent with previously found relationships with ROS in grasslands (Cheney and Gould 1995), though the effect had not been reproduced in shrubland fuels (Marsden-Smedley and Catchpole 1995; Catchpole et al. 1998).

That RH was most consistently and strongly correlated with burned area per day across all subsets is consistent with previous chaparral fuel studies on the probability of fire spread in laboratory fuelbeds (Zhou et al. 2005) and on ROS in prescribed burns (Lindenmuth and Davis 1973). It was somewhat unexpected that RH would be a stronger predictor than FM because the body of work showing FM as a predictor of ROS (Lindenmuth and Davis 1973; McCaw 1997; Fernandes 2001; Baeza et al. 2002) and probability of spread (Weise et al. 2005; Anderson and Anderson 2010) in chaparral-type fuels is more robust than that of RH. However, RAWS FM measurements are not necessarily comparable; these studies use laboratory measures of FM of fine fuels, litter or live 10-h fuels as opposed to the automated 10-h dead FM measurements reported by RAWS units.

Numerous laboratory and prescribed burn studies corroborate that WS is a key driver of ROS in chaparral fuels (McCaw 1997; Catchpole et al. 1998; Fernandes 2001; Bilgili and Saglam 2003; Koo et al. 2005). Some have successfully modelled ROS in chaparral with WS as the only input parameter (Clark et al. 2004; Zhou et al. 2005), whereas others suggest a less dominant role of WS that is conditional on or secondary to favourable FM (Lindenmuth and Davis 1973; Baeza et al. 2002; Morvan et al. 2002) or fuel loading (Fendell and Wolff 2001; Bilgili and Saglam 2003) conditions. Our results regarding the role of WS were constrained by missing data in many of the assigned RAWS units (just over half burned areas had RAWS peak WS data). Further, the ability of RAWS WS measurements to reliably capture SA-related wind gusts is questionable due to the once-hourly reporting schedule as well as the positioning of some units near obstructions (Fovell 2012). Nevertheless, peak WS is shown here to be a key predictor of wildfire progression, particularly on the first day of fire events, among large (>500 ha) burned areas, and on SA days (Table 4). The strong association with the first-day-of-fire-event subset is explained by the predominant occurrence of SA events on those first days: SA events (by any definition) occurred on 43% of first-day burned areas and only on 18% of burned areas otherwise. The weakened correlation of WS with burned area after the first day is likely related to changing wind direction: mean wind direction for first-day-of-fire-event was 353° whereas after the first day, wind direction averaged 234°. This directional shift by 119° suggests that after the first day, winds were blowing perpendicular to or into fuels that had already been burned, weakening the correlation.

The temporal position of a burn within a fire event has not previously been studied since large samples of fire progression data have not previously been available, but it had a large influence on predicting daily fire progression. As previously discussed, the first-day subset captures SA-related effects on PD, HD, WS and wind direction. The last-day effect, particularly the relatively strong positive correlation with PD and HD, is rather interesting. One possible explanation is that aggressive containment efforts disrupt natural fire progression rates; fires in high population density areas may be closer to their peak rate of burning on the day they are contained than fires in less populated areas that instead undergo a more gradual natural decline.

In summary, large-scale data on the predictors of fire progression such as those produced in this study can be most advantageously used to validate results from smaller-scale test data (Fendell and Wolff 2001). Given the paucity of published in situ data on the predictors of wildfire spread, particularly regarding extreme fire events, we anticipate that our predictor correlation data will be useful in the ongoing development, validation and refinement of both fire spread models as well as fire weather danger rating systems. A continued need exists for large-scale in situ data on predictors of wildfire progression, particularly in other regions and fuel types.

Acknowledgements

This study was supported by a grant from the National Institute of Environmental Health Sciences, one of the National Institutes of Health. We thank L. Giglio (University of Maryland) for access to preliminary versions of the MODIS MCD64A1 dataset. This study was made possible in part due to the data made available by the governmental agencies, commercial firms and educational institutions participating in MesoWest. We thank the anonymous reviewers of a previous submitted version of this paper.

References

  1. Anderson SAJ, Anderson WR (2010) Ignition and fire spread thresholds in gorse (Ulex europaeus). International Journal of Wildland Fire 19, 589–598. doi: 10.1071/WF09008 [DOI] [Google Scholar]
  2. Baeza M, De Lu’ıs M, Ravent’os J, Escarr’e A (2002) Factors influencing fire behaviour in shrublands of different stand ages and the implications for using prescribed burning to reduce wildfire risk. Journal of Environmental Management 65, 199–208. doi: 10.1006/JEMA.2002.0545 [DOI] [PubMed] [Google Scholar]
  3. Bilgili E, Saglam B (2003) Fire behavior in maquis fuels in Turkey. Forest Ecology and Management 184, 201–207. doi: 10.1016/S0378-1127(03)00208-1 [DOI] [Google Scholar]
  4. Burrows N, Ward B, Robinson A (1991) Fire behavior in spinifex fuels on the Gibson Desert Nature Reserve, Western Australia. Journal of Arid Environments 20, 189–204. [Google Scholar]
  5. Catchpole W, Bradstock R, Choate J, Fogarty L, Gellie N, McArthy G, McCaw L, Marsden-Smedley J, Pearce G (1998) Co-operative development of equations for heathland fire behaviour. In ‘Proceedings of the 3rd International Conference on Forest Fire Research and 14th Conference on Fire and Forest Meteorology’, 16–20 November 1998, Luso, Portugal. (Ed. Viegas DX) pp. 631–645. (ADAI, University of Coimbra: Coimbra, Portugal: ) [Google Scholar]
  6. Cheney NP, Gould JS (1995) Fire growth in grassland fuels. International Journal of Wildland Fire 5, 237–247. doi: 10.1071/WF9950237 [DOI] [Google Scholar]
  7. Clark TL, Coen JL, Latham D (2004) Description of a coupled atmosphere–fire model. International Journal of Wildland Fire 13, 49–63. doi: 10.1071/WF03043 [DOI] [Google Scholar]
  8. Coen JL, Douglas CC (2010) Computational modeling of large wildfires: a roadmap. In ‘Ninth International Symposium on Distributed Computing and Applications to Business, Engineering and Science’, 10–12 August 2010, Hong Kong, China, pp. 113–117. (IEEE Xplore Digital Library: Piscataway, NJ: ) [Google Scholar]
  9. Conil S, Hall A (2006) Local regimes of atmospheric variability: a case study of southern California. Journal of Climate 19, 4308–4325. doi: 10.1175/JCLI3837.1 [DOI] [Google Scholar]
  10. Davis FW, Michaelsen J (1995) Sensitivity of fire regime in chaparral ecosystems to climate change. In ‘Global Change and Mediterranean-type Ecosystems’. (Eds Moreno J, Oechel WC) pp. 435–456. (Springer-Verlag: New York: ) [Google Scholar]
  11. Edinger JG, Helvey RA, Baumhefner D (1964) Surface wind patterns in the Los Angeles basin during ‘Santa Ana’ conditions. USDA Forest Service, Part 1 of USDA Forest Service Research Project 2606, Pacific Southwest Forest and Range Experiment Station. (Los Angeles, CA: ) [Google Scholar]
  12. Eidenshink J, Schwind B, Brewer K, Zhu Z, Quayle B, Howard S (2007) A project for monitoring trends in burn severity. Fire Ecology 3, 3–21. doi: 10.4996/FIREECOLOGY.0301003 [DOI] [Google Scholar]
  13. Fendell FE, Wolff MF (2001) Wind-aided fire spread. In ‘Forest Fires: Behavior and Ecological Effects’. (Eds Johnson EA, Miyanishi K) pp. 171–224. (Academic Press: San Diego, CA: ) [Google Scholar]
  14. Fernandes PM (2001) Fire spread prediction in shrub fuels in Portugal. Forest Ecology and Management 144, 67–74. [Google Scholar]
  15. Fernandes PM, Botelho HS, Lourerio C (2002) Models for the sustained ignition and behavior of low-to-moderately intense fires in maritime pine stands. In ‘Proceedings of IV International Conference on Forest Fire Research and 2002 Wildland Fire Safety Summit’, 19–21 June 1978, Lake Tahoe, CA. (Ed Viegas DX) (Millpress: Rotterdam, Netherlands: ) [Google Scholar]
  16. Fosberg MA (1978) Weather in wildland fire management: the fire weather index. In ‘Proceedings of the Conference on Sierra Nevada Meteorology’, 19–21 June 1978, South Lake Tahoe, NV. pp. 1–4. [Google Scholar]
  17. Fovell RG (2012) Downslope windstorms of San Diego county: sensitivity to resolution and model physics. In ‘13th WRF Users’ Workshop’, 26–29 June 2012. (National Center for Atmospheric Research: Boulder, CO: ) Available at http://www2.mmm.ucar.edu/wrf/users/workshops/WS2012/WorkshopPapers.php [Verified 6 October 2014 [Google Scholar]
  18. Gelman A (2008) Scaling regression inputs by dividing by two standard deviations. Statistics in Medicine 27, 2865–2873. doi: 10.1002/SIM.3107 [DOI] [PubMed] [Google Scholar]
  19. Giglio L, Descloitres J, Justice CO, Kaufman YJ (2003) An enhanced contextual fire detection algorithm for MODIS. Remote Sensing of Environment 87, 273–282. doi: 10.1016/S0034-4257(03)00184-6 [DOI] [Google Scholar]
  20. Giglio L, Loboda T, Roy DP, Quayle B, Justice CO (2009) An active-fire based burned area mapping algorithm for the MODIS sensor. Remote Sensing of Environment 113, 408–420. doi: 10.1016/J.RSE.2008.10.006 [DOI] [Google Scholar]
  21. Hughes M, Hall A (2010) Local and synoptic mechanisms causing southern California’s Santa Ana winds. Climate Dynamics 34, 847–857. doi: 10.1007/S00382-009-0650-4 [DOI] [Google Scholar]
  22. Keeley JE, Fotheringham CJ (2001) Historic fire regime in southern California shrublands. Conservation Biology 15, 1536–1548. doi: 10.1046/J.1523-1739.2001.00097.X [DOI] [Google Scholar]
  23. Keeley JE, Zedler PH (2009) Large, high-intensity fire events in southern California shrublands: debunking the fine-grain age patch model. Ecological Applications 19, 69–94. doi: 10.1890/08-0281.1 [DOI] [PubMed] [Google Scholar]
  24. Keeley JE, Safford H, Fotheringham CJ, Franklin J, Moritz M (2009) The 2007 southern California wildfires: lessons in complexity. Journal of Forestry 107, 287–296. [Google Scholar]
  25. Koo E, Pagni P, Stephens S, Huff J, Woycheese J, Weise DR (2005) A simple physical model for forest fire spread rate. Fire Safety Science 8, 851–862. doi: 10.3801/IAFSS.FSS.8-851 [DOI] [Google Scholar]
  26. Lindenmuth AW, Davis JR (1973) Predicting fire spread in Arizona’s oak chaparral. USDA Forest Service, Rocky Mountain Forest and Range Experiment Station, RM-101. (Fort Collins, CO; ) [Google Scholar]
  27. Linn RR, Winterkamp JL, Weise DR, Edminster C (2010) A numerical study of slope and fuel structure effects on coupled wildfire behaviour. International Journal of Wildland Fire 19, 179–201. doi: 10.1071/WF07120 [DOI] [Google Scholar]
  28. Loboda TV, Csiszar IA (2007) Reconstruction of fire spread within wildland fire events in Northern Eurasia from the MODIS active fire product. Global and Planetary Change 56, 258–273. doi: 10.1016/J.GLOPLACHA.2006.07.015 [DOI] [Google Scholar]
  29. Marsden-Smedley JB, Catchpole WR (1995) Fire behaviour modeling in Tasmanian buttongrass moorlands: I. fuel characteristics. International Journal of Wildland Fire 5, 203–214. doi: 10.1071/WF9950203 [DOI] [Google Scholar]
  30. McCaw L (1997) Predicting fire spread in Western Australian mallee–heath shrubland. PhD thesis, University of New South Wales, Canberra. [Google Scholar]
  31. Miller NL, Schlegel NJ (2006) Climate change projected fire weather sensitivity: California Santa Ana wind occurrence. Geophysical Research Letters 33, L15711. doi: 10.1029/2006GL025808 [DOI] [Google Scholar]
  32. Moritz MA (1997) Analyzing extreme disturbance events: fire in Los Padres National Forest. Ecological Applications 7, 1252–1262. doi: 10.1890/1051-0761(1997)007[1252:AEDEFI]2.0.CO;2 [DOI] [Google Scholar]
  33. Moritz MA, Moody TJ, Krawchuk MA, Hughes M, Hall A (2010) Spatial variation in extreme winds predict large wildfire locations in chaparral ecosystems. Geophysical Research Letters 37, L04801. doi: 10.1029/2009GL041735 [DOI] [Google Scholar]
  34. Morvan D, Tauleigne V, Dupuy JL (2002) Wind effects on wildfire propagation through a Mediterranean shrub. In ‘Forest Fire Research and Wildland Fire Safety: Proceedings of IVth International Conference on Forest Fire Research/2002 Wildland Fire Safety Summit’, 18–23 November 2002, Coimbra, Portugal. (Ed. Viegas DX) pp. 116–130 (Millpress: Rotterdam, Netherlands: ) [Google Scholar]
  35. NWCG (2011) Historical Incident ICS-209 Reports. (National Wildlife Coordinating Group; ) Available at http://fam.nwcg.gov/fam-web/hist_209/report_list_209 [Verified 23 September 2014] [Google Scholar]
  36. Ottmar RD, Sandberg DV, Riccardi CL, Prichard SJ (2007) An overview of the Fuel Characteristic Classification System – quantifying, classifying, and creating fuelbeds for resource planning. Canadian Journal of Forest Research 37, 2383–2393. doi: 10.1139/X07-077 [DOI] [Google Scholar]
  37. Peterson SH, Moritz MA, Morais ME, Dennison PE, Carlso JM (2011) Modelling long-term fire regimes of southern California shrublands. International Journal of Wildland Fire 20, 1–16. doi: 10.1071/WF09102 [DOI] [Google Scholar]
  38. Price OF, Bradstock RA, Keeley JA, Syphard AD (2012) The impact of antecedent fire area on burned area in southern California coastal ecosystems. Journal of Environmental Management 113, 301–307. doi: 10.1016/J.JENVMAN.2012.08.042 [DOI] [PubMed] [Google Scholar]
  39. Raffuse SM, Sullivan DC, Larkin NK, Gilliland EK, Chinkin LR, Solomon R, Pace TG (2008) Development and sensitivity analysis of wildland fire emission inventories for 2002–2006. In ‘17th International Emissions Inventory Conference’, 2–5 June 2008, Portland, OR. pp. 2–5. [Google Scholar]
  40. Raphael MN (2003) The Santa Ana winds of California. Earth Interactions 7, 1–13. doi: [DOI] [Google Scholar]
  41. Richardson RT (1973) The continental contribution to the climate of southern California. PhD thesis, University of Oregon, Eugene. [Google Scholar]
  42. Rothermel RC (1972) A mathematical model for predicting fire spread in wildland fuels. USDA Forest Service, Intermountain Forest and Range Experiment Station, INT-115. [Google Scholar]
  43. Schroeder MJ, Glovinsky M, Henricks VF, Hood FC, Hull MK (1964) Synoptic weather types associated with critical fire. USDA Forest Service, Pacific Southwest Range and Experiment Station. (Berkeley, CA: ). [Google Scholar]
  44. Sergius LA, Huntoon JK (1956) An objective method for forecasting the Santa Ana. In ‘Forecaster’s Handbook’. (US Fleet Weather Central: San Diego, CA: ). [Google Scholar]
  45. Sullivan AL (2009) Wildland surface fire spread modeling, 1990–2007. International Journal of Wildland Fire 18, 369–386. doi: 10.1071/WF06142 [DOI] [Google Scholar]
  46. Turner MG, Hargrove WW, Gardner RH, Romme WH (1994) Effects of fire on landscape heterogeneity in Yellowstone National Park, Wyoming. Journal of Vegetation Science 5, 731–742. doi: 10.2307/3235886 [DOI] [Google Scholar]
  47. US CENSUS Bureau (2002) Census 2000 Summary File 3 – California. (US Census Bureau: Washington, DC: ). [Google Scholar]
  48. US Geological Survey (1999) USGS National Elevation Dataset. (EROSData Center: Sioux Falls, SD: ). [Google Scholar]
  49. US Geological Survey (2010) LANDFIRE Refresh 2008 Fuel Characterization Classification System Fuelbeds. Available at http://landfire.cr.usgs.gov/viewer/ [Verified 23 September 2014]
  50. Weise DR, Zhou X, Sun L, Mahalingam S (2005) Fire spread in chaparral – ‘go or no-go?’. International Journal of Wildland Fire 14, 99–106. doi: 10.1071/WF04049 [DOI] [Google Scholar]
  51. Westerling AL, Cayan DR, Brown TJ, Hall BL, Riddle LG (2004) Climate, Santa Ana winds and autumn wildfires in southern California. American Geophysical Union EOS Transactions 85, 289–300. doi: 10.1029/2004EO310001 [DOI] [Google Scholar]
  52. Zhou X, Mahalingam S, Weise D (2005) Modelling of marginal burning state of fire spread in live chaparral shrub fuel bed. Combustion and Flame 143, 183–198. doi: 10.1016/J.COMBUSTFLAME.2005.05.013 [DOI] [Google Scholar]

RESOURCES