1. Introduction
Wildfires and other forest disturbances such as insect outbreaks and drought cause widespread forest mortality across the western United States [
1,
2,
3]. Forest disturbance has increased in recent years, and is projected to continue to increase in coming decades, in part due to climate change [
4,
5,
6,
7,
8]. Many aspects of wildfire behavior and effects are greatly influenced by forest fuels [
9,
10], which are shaped by prior disturbances such as insect epidemics, drought, fire, and disease. Spatially-explicit estimates of forest fuels and the effect of insect-caused tree mortality on fuels can serve as valuable information for forest managers and researchers in their efforts to predict and mitigate the effects of future fire.
Forest canopy fuels are commonly characterized with several parameters important to predicting fire behavior [
11]. Available canopy fuel (ACF) is defined as fuel capable of burning in a crown fire, and includes foliage and small branches. Canopy bulk density (CBD) describes the density of ACF in the canopy, or the amount of ACF per unit canopy volume. Canopy base height (CBH) and canopy height (CH) refer to the minimum and maximum heights at which canopy bulk density reaches a critical mass capable of supporting fire [
11]. Common surface fuel measurements include litter, defined as loose plant debris on the forest floor such as branches, twigs, and fallen needles not altered by decomposition, and duff, defined as the layer of decomposing organic material below litter and above mineral soil [
12]. Downed dead wood (DWD) surface fuels are often defined by how quickly they will reach a new moisture equilibrium, which is a function of their surface area/volume ratio. In the U.S., DWD fuels include 1-, 10-, 100-, and 1000-h fuels that correspond to woody debris with diameter classes of 0–0.6, 0.6–2.5, 2.5–7.6, and 7.6–20.3 cm, respectively [
9].
Airborne light detection and ranging (lidar) actively senses the three-dimensional structure of underlying terrain and vegetation, making it an ideal technology for spatially-explicit measurements of forest characteristics such as fuels [
13,
14]. Multispectral instruments record spectral information that lidar data lack and therefore complement information gained from lidar. Repeated lidar measurements are rare [
15] and consequently, multispectral satellite imagery is more relevant than lidar for forest monitoring because an orbiting satellite can provide repeated coverage. For instance, individual Landsat satellites (Thematic Mapper (TM), Enhanced Thematic Mapper Plus (ETM+), or Operational Land Imager (OLI)) gather imagery for a given location on earth every sixteen days; this interval decreases to eight days if multiple Landsat satellites are used. With the opening of the Landsat archive, Landsat time series approaches that make use of the entire Landsat TM, ETM+, and OLI archive (1984–present) have proven to be very useful for forest measurement, especially for detection and quantification of forest disturbance [
16,
17,
18].
Previous studies have shown how forest canopy fuels can be characterized or predicted from multispectral [
19,
20], airborne lidar [
21,
22,
23,
24,
25,
26,
27,
28,
29,
30,
31], spaceborne lidar [
32,
33], and the combination of airborne lidar and multispectral data [
34,
35,
36,
37]. Fine-scale characterizations of canopy and surface fuels using terrestrial lidar or photogrammetry have also been performed [
38,
39,
40,
41,
42]. Studies characterizing or predicting surface fuels from airborne lidar or satellite data are less numerous [
31,
36,
43,
44,
45,
46,
47,
48,
49,
50]; to our knowledge, no previous study has predicted forest surface fuels from a combination of lidar and Landsat time series data.
The effect that bark beetle-caused forest mortality has on forest fuels and subsequent fire behavior is somewhat controversial, although general agreement exists in many aspects [
51]. Time since disturbance is crucial in determining bark beetle mortality impacts on fuels and fire behavior. Beetle-killed trees generally retain needles for 2–4 years after attack [
52,
53]. During this time, known as the ‘red’ phase of attack, ACF, CBD and CBH remain unchanged, with reduced foliar moisture possibly causing an increase in crown fire potential [
54,
55]. However, few studies have documented the effects of beetle-caused tree mortality on fire behavior during this phase [
51] and some of the studies in question report contradictory results [
56,
57]. As beetle-killed trees lose needles, entering the ‘gray’ phase, ACF and CBD decrease, thereby reducing crown fire potential [
56,
58,
59,
60,
61], while concomitantly increasing fine surface fuel loads and corresponding properties of surface fire [
60,
62,
63,
64]; coarse surface fuels remain unchanged during the gray stage [
56,
59,
61,
63]. The response of CBH during the gray phase of attack is unclear, as results have varied across different studies [
51,
54,
56,
58,
59,
60,
61]. Coarse surface fuels increase as beetle-killed snags fall [
55,
56,
61,
62,
63], which occurs gradually beginning 3–5 years after attack [
65,
66,
67]. More studies on how bark beetle-caused tree mortality affects fuels are needed.
Several studies have predicted surface fuels from remote sensing data; however, to our knowledge, no previous study has used remote sensing data to predict both canopy and surface fuels in a predominantly lodgepole pine forest affected by extensive bark beetle mortality; and no previous study has examined the relationship between bark beetle-caused tree mortality and fuels across a landscape. Here, we (1) demonstrated the prediction of canopy and surface fuels from lidar and Landsat time series data using random forest (RF) modeling; (2) applied RF models to produce fuel maps; and (3) compared fuel maps to an independent assessment of forest mortality to assess relationships between fuels and mortality severity. Our results showed that canopy and surface fuels can be predicted from lidar and Landsat time series data with moderate accuracy. Such predictions could be used to improve fuel maps, fire behavior modeling, and programs like LANDFIRE. We also showed that canopy fuels decreased 2–6 years after a severe bark beetle outbreak, and that many aspects of fuels did not appear to be strongly affected by variation in the severity of tree mortality within our study area.
3. Results
Across the 119 8-m-radius plots, available canopy fuel averaged 7.3 Mg ha
−1 and canopy bulk density averaged 0.09 kg m
−3 (
Table 3). Canopy base height and canopy height averaged 2.5 and 16.6 m, respectively. Litter and duff, 1–100-h surface fuel, 1000-h surface fuel, and total surface fuel averaged 34.3, 9.4, 18.0, and 61.5 Mg ha
−1, respectively.
RF models explained 28–70% of the variation in canopy fuel variables (
Table 4); including Landsat time series variables in the models increased the percent of variance explained by 2–7%. Surface fuel variables were predicted less accurately, RF models explaining 16–32% of variance. The addition of Landsat time series variables to the models increased the percent of variance explained in surface fuels by 2–8%.
The greater predictive capacity of lidar variables relative to satellite variables is not surprising, as lidar directly senses the three-dimensional structure of the canopy and to a lesser extent, the subcanopy and forest floor. In addition, the lidar-generated variables have a higher spatial resolution (generated from multiple returns per m2) compared to the multispectral Landsat data (30-m spatial resolution), adding to the increased predictive power of the lidar data. Surface fuel variables were predicted with less accuracy than canopy variables because both airborne lidar and satellite image sensors are more sensitive to fuel structure in the forest overstory canopy than in the forest floor. However, lidar and satellite sensors did provide some information about the forest floor, collectively explaining 24–32% of variation in surface fuel variables.
Of the lidar-derived metrics, canopy density variables, especially the percentage of returns >1.37 m in height (DNS) and the proportion of returns >10 and <20 m in height (D05 PROP), were important predictors of canopy fuel variables (
Table 5). The skewness of canopy returns heights (SKE) was the most important predictor of CBH, and maximum canopy return height (MAX) was the most important predictor of CH. SKE and MAX are logical selections for describing CBH and CH, respectively; SKE is a measure of canopy profile shape from which CBH is derived (
Figure 2), and MAX closely approximates CH. Elevation above sea level (ELEV) helped explain variation in ACF and CH; both were positively correlated with ELEV. Only three LandTrendr variables were selected as important predictors of canopy fuel variables; pre-disturbance Landsat tasseled cap wetness (GDpreTCW), a measure of vegetation moisture content, was an important predictor of both ACF and CBD. Percent cover of bark beetle-caused tree mortality (MORT%) was selected as an important predictor of–and negatively correlated with–ACF. Both MORT% and the rate of bark beetle-caused tree mortality (MORTrate) helped explain variation in CBH.
Similar to canopy fuel variables, lidar density variables, both lower strata variables and upper strata variables were important predictors of surface fuels (
Table 5). The selection of both lower and upper density strata explanatory variables in the prediction of surface fuels suggests that lidar was sensitive to surface fuels both directly and indirectly. In particular, the fact that the skewness of returns >0.15 and <0.5 m in height (D01 SKE) was chosen as an important predictor of surface fuels is evidence that lidar directly interacted with surface fuels, whereas the selection of an upper canopy variable, the 75th percentile of canopy return height (P75), as the most important predictor of 1–100-h and total surface fuels shows that this variable described surface fuels indirectly through association with canopy conditions.
Landsat time series variables were chosen as important predictors of surface fuels more frequently than they were chosen for canopy fuel variables, likely because lower performance of lidar variables left substantial variation for the satellite variables to explain. The most important predictor of litter and duff was post-disturbance canopy cover (LDpost.val), which presumably represents an indirect measure of litter and duff as they accumulate below dead trees. Pre-disturbance Landsat tasseled cap greenness (LDpreTCG), an indicator of vegetation density [
96], was a similar indirect measure of fine surface fuels. Pre-disturbance Landsat tasseled cap brightness (LDpreTCB), sensitive to soil characteristics, might have been a more direct indicator of fine surface fuels. Pre-disturbance Landsat tasseled cap wetness (GDpreTCW) and the duration of the greatest disturbance (GDdur) were important predictors and positively correlated with 1000-h and total surface fuel loads. GDdur described both the recent bark beetle disturbance and previous disturbance, which LandTrendr detected across much of the study area (
Figure 4). MORT% was an important predictor of 1000-h surface fuels, the relationship between 1000-h surface fuels and MORT% being negative (
ρ = −0.4).
Two topographic variables, ELEV and curvature, the second derivative of the slope (CURV) [
97], were selected as important predictors of fine surface fuels. ELEV was negatively correlated with 1–100-h surface fuels. The relationship between CURV and litter/duff is unclear; however, CURV describes slope shape, whether upwardly convex (positive CURV) or upwardly concave (negative CURV), which might affect the degree to which litter and duff are subject to erosion.
Maps of fuel response variables show mortality-related patterns (
Figure 1 and
Figure 5) that were confirmed through Spearman’s correlation tests using independent ADS-derived mortality data (
Table 6). In forested areas affected by bark beetles, ACF, CBD, and CH decreased with increasing tree mortality severity, while CBH increased with increasing tree mortality severity. Though weak, most relationships were significant.
Relationships between surface fuels and mortality severity were predominantly negative, somewhat contrary to expectations. No significant relationship between litter and duff and ADS-derived mortality severity was found. A weak negative relationship between 1-h to 100-h surface fuels and mortality severity existed, as well as a moderate negative relationship between 1000-h surface fuels and mortality severity. Coarse surface fuels were likely greater in areas of less severe bark beetle mortality at the time of our sampling (2–6 years after peak epidemic mortality in the region) because these areas had experienced previous disturbance, as revealed by LandTrendr (
Figure 4), resulting in reduced forest density and lower availability of host trees for bark beetles in the 2000s epidemic.
4. Discussion
We predicted canopy and surface fuel loads with moderate accuracy in a lodgepole pine-dominated forest affected by an extensive bark beetle outbreak. When predicting these fuel metrics, lidar-derived variables explained much more variation than did multispectral Landsat variables, and Landsat variables added minor to moderate improvements to the predictive capacity of lidar-based models. Similar findings have been reported for predicting forest stand structure attributes such as basal area [
86]. We found that the main effects of the bark beetle outbreak shortly (2–6 years) after tree mortality in our study area were a decrease in canopy height and canopy fuels, with an increase in CBH. The maps we generated of these metrics could be used by land managers, policy makers, and fire fighters to assess the impacts of bark beetle disturbances on landscapes that are generally perceived as highly flammable and hazardous [
98].
We explained less variation in canopy fuels than others have in previous studies that also used lidar, and in some cases multispectral imagery, to estimate continuous canopy fuel variables in undisturbed forests. Andersen et al. [
21] explained 86%, 84%, 77%, and 98% of the variation in sqrt(ACF), ln(CBD), CBH, and CH, respectively, in a Douglas-fir forest. For the same forest type, Hermosilla et al. [
28] reported
R2 values of 79%, 67%, 78%, and 79% for ACF, CBD, CBH, and CH, respectively. In a ponderosa pine forest, Hall et al. [
22] achieved
R2 values of 83% and 80% for CBD and CBH, respectively. Popescu and Zhao [
24] and Vauhkonen [
26] explained 80% and 84% of the variation in CBH of individual trees. Erdody and Moskal [
34] included both lidar and aerial imagery explanatory variables to achieve
R2 values of 90, 87, 83, and 95% for ACF, CBD, CBH, and CH, respectively, in a mixed conifer forest. Skowronski et al. [
27] predicted ACF and CBD in a pitch pine forest with
R2 values of 71 and 83%, respectively. Using ICESat and GLAS satellite lidar data, García et al. [
32] explained 78% of the variation in CBD. Jakubowski et al. [
36] predicted CBD, CBH, and CH with
R2 values of 25, 41, and 87%, respectively, in a mixed conifer forest. In a
Pinus radiata forest in Spain, González-Ferreiro et al. [
29] predicted ACF, CBD, CBH, and CH with
R2 values of 82%, 44%, 98%, and 98%, respectively, using low-density lidar.
There are a few reasons why we explained less variation in fuels than others have in previous studies. First, previous studies have not predicted canopy fuels in lodgepole pine forests, which, in contrast to other forest types previously studied, tend to be denser, with closed canopies and relatively shorter, uniform tree heights in many stands leading to less variability available to explain in our data. Using higher density lidar that is better able to penetrate the forest canopy or adjusting for the attenuation of the lidar from the canopy to the surface [
99] might improve the sensitivity of lidar to variation in fuels, especially surface fuels, in these relatively shorter and denser lodgepole pine forests. Second, we might have explained more variation in fuels if our field plots had been larger; preliminary models that used the field plot data of Meddens et al. [
100], which had larger 400-m
2 plot extents, performed comparably to those of previous studies [
101,
102]. We chose not to use this field data, however, because it had no associated surface fuel measurements and was gathered two years previous to our lidar data in a forest that was undergoing rapid change. Third, we did not include lidar intensity although our preliminary models also showed lidar intensity (amount of energy reflected back to the sensor) variables to be important predictors that improved model performance; however, in this study we were not able to normalize intensities so intensity information was withheld from analysis. Normalization of lidar intensity values is necessary to control for the effect of differing range between the plane and location of returns, and automatic gain control (AGC) of the lidar instrument during acquisition; intensities could have been normalized if we had had access to information on the smoothed best estimate of trajectory (SBET) of the plane and AGC information from the lidar instrument.
Few previous studies have estimated surface fuels from lidar or satellite data. The percentages of variance explained by our models- 25–32% for surface fuel variables- are similar to those of previous studies. Pesonen et al. [
45] explained 61% of the variation in downed dead wood volume using airborne lidar. Using airborne SAR and optical data, Huang et al. [
46] achieved moderate correlation (
R = 0.54) between measured and predicted coarse woody debris. Meigs et al. [
47] explained 29% of the variation in coarse woody debris with LandTrendr disturbance magnitude. Using lidar and multispectral variables in a mixed conifer forest, Jakubowski et al. [
36] attained
R2 values of 32% and 48% for 1000-h and total surface fuels, respectively, and also explained 35% of the variation in fuel bed depth. Hudak et al. [
49] explained 32–44% of the variation in surface fuels in an open longleaf pine forest. Price and Gordon [
31] explained 24% of the variation in surface fuel loads in dry sclerophyll forests in Australia. Like canopy fuel variables, we found lidar intensity variables to be effective predictors of surface fuel variables in preliminary models. Future attempts to predict surface or canopy fuels from lidar might benefit from the inclusion of normalized lidar intensity information [
103].
We found that canopy fuel responses to bark beetle mortality generally conformed to expectations and previous research. Decreases in ACF and CBD caused by the loss of needles and fine branches during the gray stage are well-documented [
56,
58,
59,
60,
61]. However, to our knowledge no previous study has demonstrated the negative correlations we found between ACF, CBD and severity of tree mortality across a landscape. Previous studies are less clear about the effect that tree mortality has on CBH, with some reporting a decrease in CBH [
58,
62] or no change in CBH [
51,
56,
60]. Like Klutsch et al. [
59] and Donato et al. [
61], we found that CBH increased following mortality, and demonstrated that CBH was positively and significantly correlated with tree mortality severity across the landscape. A post-outbreak increase in CBH is logical, as a decrease in overall CBD would cause the height of the critical lower threshold of 0.012 kg m
−3 that defines CBH to increase [
61]. For the same reason, the decrease in CH with increasing mortality severity that we found is also logical, as a decrease in ACF and CBD in the upper canopy would decrease the height of the upper critical threshold that defines CH.
Relationships between surface fuels and mortality that we found should be treated with less certainty, as percent variance explained by surface fuel models was fairly low (
Table 4). Low model performance could be due not just to lack of satellite and lidar data sensitivity but also by the inherent high variability in fine surface fuel distributions on the ground, making them difficult to sample and predict [
104]. Variables selected as important for the prediction of surface fuels suggest that lidar and Landsat data measured surface fuels both directly and indirectly, i.e., surface fuels were indirectly predicted through their association with canopy characteristics [
44,
50,
105]. We found no significant relationship between litter and duff and mortality severity, and a weak negative relationship between 1–100-h surface fuels and mortality severity. Although a positive relationship between fine surface fuels and mortality severity might be expected due to bark beetle-killed trees dropping dead needles and small branches [
51,
60,
62,
63,
64], others have also found bark beetle mortality to have no significant effect on fine surface fuels in the gray stage [
56,
59,
61], and have not detected significant relationships between fine surface fuels and satellite-derived mortality measures [
47]. While trees killed in a beetle outbreak appear to contribute a large single pulse of needles and fine branches to the forest floor, this phenomenon only occurs over several years in a given area during the course of an epidemic, and may not strongly influence the distribution of fine surface fuels relative to contributions from live trees over longer periods of time.
The negative relationship between coarse surface fuels and mortality severity that we found was somewhat unexpected. Others have reported no significant increase in coarse surface fuels in the gray stage before snag fall [
56,
59,
61,
63], and Meigs et al. [
47] similarly found no significant relationship between bark beetle outbreak duration and coarse woody debris. The outbreak in our study area began only seven years previous to our field and remote sensing measurements, with most mortality occurring 2–6 years prior to measurement. Snag fall timing is variable, generally beginning 3–5 years post attack, with low initial snag fall rates that increase with time [
65,
66,
67]; however, no studies have yet measured fall rates in the 2000s mountain pine beetle epidemic in Rocky Mountain lodgepole pine forests. In our study area, most beetle-killed trees were still standing and most 1000-h fuels were from prior disturbance. Stands regenerating from or thinned by previous disturbance were likely less vulnerable to bark beetles in the 2000s epidemic and lost fewer trees to beetles, hence the negative relationship that we found between 1000-h fuels and mortality severity. Downed coarse woody debris should accumulate in our study area as beetle-killed snags fall, eventually reversing this negative relationship [
51,
55,
59,
60,
61,
62,
63].
Previous studies disagree on how bark beetle-caused tree mortality influences subsequent fire severity. Studies have reported a decrease [
106,
107], no change [
108,
109,
110,
111,
112], an increase [
113,
114], and a mixed response [
115,
116] in fire severity following bark beetle-caused tree mortality. The extent and severity of bark-beetle caused tree mortality can vary greatly at the landscape scale, e.g., [
117,
118,
119]; how this variability in severity of mortality influences fire severity is unclear [
51], although time since attack is an important factor. The decreases in ACF and CBD and the increase in CBH that we found six years after a bark beetle outbreak in our study area suggest a corresponding decrease in potential crown fire, although additional research on the relationship between beetle-caused tree mortality and fire is still needed, given the complex temporal and spatial variation in these two processes.