Introduction

Future changes in wildfire regimes depend on the combined influences of climate, land-use change, human activity, and vegetation (Fauria et al. 2011, Higuera et al. 2015). Although there is no perfect predictor for future wildfires, we present a novel method for correlating snowmelt timing to annual area burned in a particular ecoregion. Our methods build on previous studies utilizing updated remotely sensed datasets and a new snowmelt timing algorithm to compare snowmelt timing and annual area burned from 1984 to 2012 in the Middle Rockies EPA (Environmental Protection Agency) Level III Ecoregion (US EPA 2003; hereafter, Middle Rockies Ecoregion). Our model can easily be extended to any ecoregion in the United States and may be applicable internationally in regions with complete wildfire datasets.

Previous studies have addressed the relationship between early snowmelt timing and wildfire using various methods. Medler et al. (2002) examined the spatial and temporal relationships between snow cover and wildfire from 1986 to 1996 in the 11 western US states. Although their research suggested an inverse relationship between early snowmelt timing and wildfire, no strong statistical correlation was found for the 11 western states as a whole for the period of interest, or when these data were aggregated by state boundaries. This study was limited by the use of political, rather than ecological, boundaries. This is possibly because the use of political boundaries is not ecologically meaningful, and therefore the snow and wildfire interaction that one might expect to see in the montane ecosystems of the Middle Rockies was not well captured. We were interested in identifying relationships between snowmelt timing and wildfire in forested ecoregions, therefore repeating this analysis using sensible ecological boundaries and updated datasets yielded further insight. The Medler et al. (2002) study was also limited by data availability. At the time of their study, fire data were only available for 1986 to 1996. At the time of our study, we had validated fire data for 1984 to 2012, improving our ability to observe trends in spite of decadal climate oscillations. Finally, Medler et al. (2002) used acre-weeks of snow as their snow metric, which does not specifically consider the spring timing of snowmelt. We improved upon this method by identifying a day of year (DOY) for snowmelt for each cell within our study area, providing a more direct comparison between spring snowmelt timing and wildfire occurrence.

Westerling et al. (2006) compiled a comprehensive database of 1166 large wildfires (>400 ha) in the western US between 1970 and 2003 and compared it with hydroclimatic and land-surface data. As a representation for spring snowmelt timing, they used the dates of the center of mass of annual flow for snowmelt-dominated stream flow (CT) gauge records in western North America (Westerling et al. 2006). They found that the annual wildfire frequency was inversely correlated with CT across the study region, indicating that fires corresponded to earlier spring snowmelt. Their study was limited by the use of CT as an indicator of spring timing, as the stream gauge itself is limited to the boundaries of the watershed, and CT is a timing method that is difficult to estimate in real time. For our method, we used remotely sensed data, which allowed us to define our boundaries based on the spatial resolution of the data and to identify critical snowmelt thresholds in real time at the temporal resolution of the data.

Narasimhan and Stow (2010) used MODIS (Moderate Resolution Imaging Radiometer) snow cover imagery to determine the first completely snow-free days of the year using a fractional snow cover product from 2003 to 2005. Their novel approach was hampered by cloud cover in the Arctic and spring snow events. Our method employed a cloud-free dataset and controls for spring snow events with a temporally averaged melt date. Our analysis spanning 1984 to 2012 also allowed for a more robust interannual comparison of snowmelt timing. Still, the MODIS snow data products give researchers a dramatically improved spatial resolution for environmental investigation. We kept this potential in mind during our statistical analysis when we checked whether the MODIS period of record (2000 to present) showed a different relationship between snowmelt timing and wildfire from the full study period.

Semmens and Ramage (2012) investigated the relationship between snowmelt and wildfire outbreaks using the Advanced Microwave Scanning Radiometer for EOS (AMSR-E), which can detect melt and refreeze cycles at the snow surface. Their method did not consider the timing of snow cover loss, as is calculated here. Additionally, their focus on a single watershed in Alaska may not be relevant to the US Mountain West.

Building on these findings we designed a model to explore the following question: Is there a detectable relationship between snowmelt timing and annual area burned within the Middle Rockies Ecoregion using contemporary remotely sensed datasets?

Observing patterns in wildfire and climate requires extensive and long-term datasets (Higuera et al. 2015). Our research methods used the recent Monitoring Trends in Burn Severity (MTBS) fire perimeter data (Finco et al. 2012) and weekly snow cover extent provided by the Rutgers Global Snow Lab analysis of National Oceanic and Atmospheric Administration (NOAA) daily snow maps (Robinson et al. 2012) for the study period of 1984 to 2012. Our study area was the Middle Rockies Ecoregion, which encompasses the mountainous regions of Wyoming, North Dakota, Montana, and Idaho, USA (Figure 1). These regions are often referred to in other literature as the US Northern Rockies. This region is sensitive to interactions between climate and wildfire and has experienced an increase in large fire occurrences in the last several decades (Littell et al. 2009, Dennison et al. 2014, Higuera et al. 2015).

Figure 1
figure 1

Map displaying the location of the Middle Rockies Ecoregion, within the 48 conterminous United States of America. The Middle Rockies Ecoregion includes the mountain regions of Wyoming, North Dakota, Montana, and Idaho. These regions are often referred to in other literature as the US Northern Rockies. Vector data from EPA and Natural Earth (http://www.naturalearthdata.com).

Full size image

Background

The paleorecord of the past 6000 years highlights the sensitivity of the Rocky Mountain region to changes in fire regime induced by climatic oscillations (Higuera et al. 2014). In the past four decades, the average annual area burned by large fires has increased 10-fold in western US forests and is predicted to escalate further (Westerling et al. 2014). A legacy of fire suppression is one of the greatest causes of increased fire activity in recent decades, yet has not affected many portions of the Rocky Mountains (Dennison et al. 2014). The fire regime of the Middle Rocky Mountains is characterized by a long Mean Fire Return Interval (MFRI) followed by infrequent yet high-severity, stand-replacing fires (Baker 2009). Fire suppression did not drastically impact fuel loading in the Middle Rockies Ecoregion because the MFRI of the region’s forested ecosystems is greater than the period of suppression spanning the last 100 years (Baker 2009). Together these factors made our study area an ideal region to isolate and explore the complex interaction between climate, in particular snowmelt timing, and fire.

Wildfire is strongly linked to climate variability (Gedalof et al. 2005, Miller et al. 2008, Littell et al. 2009, Fauria et al. 2011, Higuera et al. 2015). Ultimately, fuel moisture content and length of fire season have the largest observed influence on annual area burned (Westerling et al. 2006, Spracklen et al. 2009, Higuera et al. 2015). Many possible climatic factors can lead to dry fuels and increased fire danger. Climatic metrics such as Palmer Drought Severity Index (PDSI), Growing Degree Days, summer and spring mean temperature, and fuel moisture metrics such as Duff Moisture Code can accurately explain upwards of 50 % of annual area burned in historical North American wildfires (Gedalof et al. 2005, Westerling et al. 2006, Miller et al. 2008, Spracklen et al. 2009, Higuera et al. 2015).

Snow cover is a first order control on an ecosystem. The presence or absence of snow has an influence on energy balance, the timing and robustness of vegetation, soil moisture, weathering, and other processes (Moore et al. 2014). As with other climatic factors, early snowmelt can lead to increased annual area burned through the mechanism of preconditioning large expanses of fuels to low moisture content and extending fire seasons (Westerling et al. 2006). Remotely sensed snowmelt timing has the potential to advise land managers months before the beginning of fire season. Whereas PDSI, summer temperatures, and fuel moisture levels explain a high percentage of fire activity, they are only available weeks to days before a fire. The snowmelt and wildfire interactions analyzed here will be of particular interest to logistics managers as they distribute resources in preparation of fire season.

The predicted future fire regime of the US Mountain West may not support the persistence of the current community of species (Westerling et al. 2011). As the snow melts earlier, fires become more likely at higher elevations and the MFRI is reduced (Westerling et al. 2006). Predictive models suggest that, as the climate warms, wildfire may become as important as direct effects of climate change in the persistence of high elevation vegetation, including regions in which wildfire is currently rare or absent (Schumacher and Bugmann 2006). Understanding the role of wildfire and snowmelt is critical in order to assess the potential impacts of climate change on North American mountain ecosystems.

Methods

For this analysis, our region of study was the Middle Rockies 6.2.10, as defined by EPA Ecoregions Level III (Figure 1). To clarify geographical discrepancies, not all studies define the regions of the Rockies using EPA Ecoregions Level III. For example, the region that Westerling et al. (2006) describe as the “Northern Rockies” overlaps with the Middle Rocky Mountain EPA Ecoregion Level III, including the Greater Yellowstone Ecosystem.

Snow data were acquired from the Rutgers University Global Snow Lab (Robinson et al. 2012). Snow presence in this dataset was mapped weekly on an 89 × 89 cell Cartesian grid laid over a polar stereographic projection of the Northern Hemisphere beginning in the 1960s (Medler et al. 2002). This NASA (National Aeronautics and Space Administration) MEaSUREs (Making Earth System Data Records for Use in Research Environments) Earth System Data Record (ESDR) contains two separate readings for snow presence: one using the weekly NOAA-NCDC (National Climatic Data Center) Northern Hemisphere Snow Cover Extent Climate Data Record (CDR), and the other from a gap-filled snow extent product derived from the Spectral Sensor Microwave-Imager (SSMI) and the Spectral Sensor Microwave Imager-Sounder (SSMIS). These values were then codified based on snow presence determined by both sensors (code = 10), CDR only (11), SSMI(S) only (12), and no snow (20). For this research, we assumed that a positive snow code from either or both sensors indicated snow presence, and made our snow or no snow determination accordingly. There were no other values reported for this region.

To determine the timing of snow loss, we considered the differences between the weekly snow presence codes. Snow loss was determined by two consecutive readings of non-snow values. Occasionally snow would return as evidenced by two consecutive snow covered values, usually as the result of a spring snow event. For these cases, we took the average of the first snowmelt DOY and the final snowmelt DOY. We reported the date of first snow loss as the first instance of the first day of 14 or more consecutive no snow values. We reported date of final snow loss as the last instance of the first date of 14 or more consecutive snow covered values. In most cases, the aforementioned two values were the same. Averaged melt date (AMD) was the average of the date of first snow loss and the date of final snow loss. This snow dataset had 20 data points that fell within the Middle Rockies region. For each year, we calculated the AMD for each point, and then averaged together to determine a mean snowmelt date for the entire region. From those values, we performed calculations for average snowmelt date and standard deviation of snowmelt dates for the period of record, and calculated a Z-value for each year on record. A negative Z-value indicated a year with an earlier snowmelt than average.

We sourced fire data from the Monitoring Trends in Burn Severity (MTBS) National Burn Area Boundaries data from 1984 to 2012 (MTBS Project 2014). In ESRI ArcGIS10.2.2, we clipped the fire area to the Middle Rocky Mountains, as defined by EPA Ecoregions Level III, resulting in 663 fires from 1984 to 2012 in our study area. From the reported burned area for each fire in the dataset, we calculated the annual area burned in the study area. We calculated the mean and standard deviation of total annual area burned for years 1984 to 2012, and calculated a Z-value for each year on record. A positive Z-value indicated a year with more fire than average.

We performed statistical analyses in R version 3.1.2 (R Project for Statistical Computing, Vienna, Austria). The Z-value of the regional snowmelt date was compared with the Z-value of total annual area burned (Figure 2). We tested our data for temporal autocorrelation using R’s ACF function and found no significant autocorrelation for the annual area burned or snowmelt data, leaving us to proceed without corrections. We then subset our dataset to evaluate different periods of record, first omitting the exceptionally large fire year of 1988 (the big Yellowstone fires), then the period of record since the MODIS instruments became operational on the Aqua and Terra satellites (2000 to 2012). We calculated a Spearman’s rank correlation between the scaled Z-values for annual area burned versus snowmelt timing for each period of interest.

Figure 2
figure 2

Comparison of snowmelt anomaly and total area burned for the subsequent fire season in the Middle Rockies Ecoregion. The largest fire year (1988) burned when the snow melted slightly later than average, whereas the major fire events of 2000, 2003, 2006, and 2012 all burned on years with an early snowmelt.

Full size image

Results

Visual comparison of total annual area burned versus snowmelt Z-values (Figure 2), and scatterplots of the Z-values for the different subsets (Figure 3), suggests that large fire years coincide with early snowmelt. Fire activity generally increased from 1999 to 2012, coinciding with negative snow anomalies, with a relative absence of fire during the later-melting 1984 to 1998 years (with the exception of 1988).

Figure 3
figure 3

Scatterplot of snowmelt date anomaly (Z-value snow coverage) and total area burned anomaly (Z-value total area burned) from 1984 to 2012. The black triangle symbolizes 1988, and years 2001 to 2012 (corresponding to MODIS period of record) are symbolized as black circles. All other years are open circles. Year 1988 is the only point in the positive-positive quadrant—the only year on record with above-average area burned following a later than average snowmelt. Years 2000 to 2012 contain all of the other above-average fire years for this period.

Full size image

All correlation tests returned negative relationships between snowmelt Z-values and annual area burned Z-values. The full period of record (p27 = −0.331, P-value = 0.0797) and the subset omitting 1988 (p26 = −0.374, P-value = 0.0497) both showed a strong relationship between snowmelt timing and wildfire occurrence. The subset corresponding to MODIS data availability (p11 = −0.104, P-value = 0.737) showed little to no relationship between these two variables.

Discussion

Our results revealed an inverse relationship between snowmelt anomaly and total annual area burned in the Middle Rockies Ecoregion from 1984 to 2012. An earlier spring snowmelt was correlated with above-average annual area burned. These results support the findings of others (Medler et al. 2002, Westerling et al. 2006, Semmens and Ramage 2012). The implications of our findings are especially critical because the average annual area burned in the western US is projected to increase further, and years without fires are predicted to be increasingly rare (Westerling et al. 2011). Additionally, as the climate changes in the US Mountain West, snow is expected to melt progressively earlier in the season (Stewart 2009). Our findings suggest that snowmelt occurring earlier in the coming decades will exacerbate wildfire risk. If this relationship continues to persist, we can expect an increase of wildfire annual area burned in the twenty-first century.

The link between snowmelt and annual area burned is apparent for the large increase in fire activity from 2000 to 2012, which coincides with a decade of early snowmelt (Figure 2). Indeed, aside from 1988, all above-average fire years occur within the 2000 to 2012 subset (Figure 3). Examination of the scatterplot (Figure 3) reveals a clear pattern that the majority of above-average fire years follow an early snowmelt anomaly, similar to findings in Medler et al. (2002). We can see in the full period of record that there is only one point (1988) that experiences a higher than average fire season following a positive snowmelt anomaly (later than average snowmelt date). All other above-average fire years follow a negative snowmelt anomaly (earlier than normal melt). There are many years with a low total annual area burned, regardless of the snowmelt anomaly. However, the years with high fire totals almost exclusively follow early snowmelt anomalies. This indicates that late snowmelt anomaly has a suppressive effect on wildfires, whereas early snowmelt anomalies are one factor in many that contribute to large fire years.

Early snowmelt is one way to achieve low fuel moisture and extended fire seasons leading to increased annual area burned, yet it is certainly not the only mechanism. Climate indirectly affects wildfire occurrence, through control over fuel moisture (Gedalof et al. 2005, Miller et al. 2008, Littell et al. 2009, Fauria et al. 2011, Higuera et al. 2015). Dry fuel loads with adequate ignitions lead to wildfires. Fuel moisture is influenced by a variety of ecological and climatic processes including high summer temperatures, drought, and snowmelt timing. Of all the factors that lead to low fuel moisture, snowmelt timing is one of the earliest predictors, occurring several months before the fire season begins. Drought indexes, fuel moisture metrics, and spring-summer temperatures may be more highly correlated with annual area burned (Gedalof et al. 2005, Westerling et al. 2006, Miller et al. 2008, Spracklen et al. 2009, Higuera et al. 2015), but are temporally concurrent with wildfire season. Our model of classifying early snowmelt timing using Rutgers Global Snow Lab analysis of NOAA daily snow maps may provide months of foresight for wildland fire managers.

While climate variables such as PDSI and air temperature have a high explanatory power for wildfire (~50 %), snowmelt timing explains very little of the variability in annual area burned (~14 %). This is not a surprise, because snowmelt occurs months before the typical fire season, and there are many factors that can influence fire conditions in the interim. Nonetheless, fire managers should include snowmelt timing in any predictive analysis as snowmelt can directly influence fuel moisture levels, which serve as one factor in many that lead to large wildfire years.

We decided a priori that the year 1988 was an outlier as it clearly had the largest annual area burned and a slightly later than average snowmelt (Figure 2, Figure 3). The majority of the fires that make up the total acreage in our study area in 1988 occurred in the Greater Yellowstone Ecosystem, burning over 250 000 ha in the surrounding areas. The large 1988 fires can be explained by many factors aside from snowmelt timing. The primary consensus of why this fire complex became so large was the combination of unusual weather conditions characterized by prolonged summer drought and extreme wind coupled with the vast extent of continuous lodgepole pine (Pinus contorta Douglas ex Loudon) forest (Romme and Despain 1989, Turner et al. 1994, Higuera et al. 2010). Historical evidence suggests that similar large fire events have occurred in Yellowstone on the order of 150- to 300-year intervals, occurring circa 1240, 1540, and 1700 (Higuera et al. 2010). The extensive fires in 1988 were caused by interactions of climate and fuel conditions, and correspond to a natural historical MFRI of 150 yr to 300 yr (Romme and Despain 1989, Turner et al. 1994, Higuera et al. 2010).

Early snowmelt is correlated with above-average wildfire occurrence, but it is not the only factor. When looking at the complete data set, the Spearman’s rank correlation results suggest that late snowmelt’s suppression of the following fire season’s annual area burned is unlikely to be random. Omitting the year 1988 shows an even stronger inverse relationship for all but the most extreme fire year. Removing 1988 is justified because this well-studied complex was caused by extreme weather following snowmelt, and by omitting this data point we are better able to see the influence of periodic fires rather than widespread stand-replacing fires. Regardless of the cause of the 1988 fires, the fire regimes of this region are shifting due to earlier snowmelt and longer fire seasons, which is predicted to dramatically alter the ecosystem by the mid-twenty-first century (Westerling et al. 2011).

Further Investigations

Future research may seek to improve upon this analysis by using an updated MTBS dataset or higher-resolution imagery. The National Snow and Ice Data Center’s MOD10A2 data-set has recently become popular (Hall et al. 2006). However, with its period of record extending back only to 2000, analyzing these higher-resolution data at a regional scale over this short period may fail to identify the relationships at play. Indeed, analysis of the subset from 2000 to 2012 yields no statistically significant relationship between snowmelt anomaly and wildfire. Improvements upon the snowmelt timing logic may be helpful in investigating the influence of spring snow storms following a lengthy period of no snow.

Conclusions

We have presented a novel method of comparing snowmelt data and fire occurrence using fine-scale resolution NOAA snow and MTBS fire data. We have improved upon existing methods by combining remotely sensed data with an algorithm that temporally averages snowmelt timing, better identifying the DOY when snow leaves the landscape. The fire regimes of various ecosystems respond differently to climatic conditions (Littell et al. 2009). Our model allows for simple comparison of snowmelt timing and annual area burned in a region of interest, and could easily be adapted across the continent or anywhere else that has a high-quality fire dataset. Further investigation could use our model on more EPA ecoregions, or improve upon the resolution of the snow extent data. Our findings bolster the results of previous studies, and further elucidate early spring snowmelt as a correlate of increased annual area burned. As snow continues to melt earlier each decade as a result of climate change, understanding this system may become increasingly important for wildfire management. Our model provides a novel tool for making general predictions about the upcoming fire season months in advance, and warrants extensive further investigation.