1. Introduction
The balsam woolly adelgid (Hemiptera:
Adelges picea Ratzeburg; hereafter BWA) is an introduced insect that affects North American true firs (
Abies spp.), causing tree damage and mortality among all age classes [
1]. Native to Europe, BWA entered North America in the early 1900s via separate introductions on the east [
2] and west coasts [
3]. In the western United States, BWA has slowly invaded true fir forests from northern California [
3] to Oregon [
4], Washington [
5], Idaho in 1983 [
6], Montana in 2007 [
7], and as of 2017, Utah [
7]. The relative threat of BWA to true fir forests varies geographically, as the severity of impacts can vary by host species, site, stand age and conditions, and local climate [
8,
9,
10,
11]. In Utah, the threat posed by BWA is high given the abundance and ecological value of a particularly susceptible host, subalpine fir (
Abies lasiocarpa (Hook.) Nutt.). Subalpine fir is the sixth-most abundant species in the state by basal area, according to FIA data [
12,
13]. It is a late-seral, shade-tolerant species often found dominating poor sites at high altitudes, occupying a unique and important niche among Utah’s forests as a wildlife habitat and acting as a critical carbon sink. Accordingly, the novel invasion of BWA into the state has the potential for devastating effects. A second true fir species in Utah, white fir (
Abies concolor (Lindley ex Hildebrand) Gordon), can be infested by BWA but is far less susceptible to severe impacts and mortality [
14].
The life history, dispersal, and cryptic nature of BWA create unique challenges for the detection and management of this pest. BWAs are microscopic (<1 mm) in size and parthenogenic, producing between two and four generations per growing season in the northwest USA [
15]. The mobile stage in BWA’s life cycle is the first instar “crawler” when BWAs first hatch from the egg and either settle on the same tree as their parent or are passively dispersed via wind or phoretic movement on birds [
16]. After selecting a feeding site on the host and inserting their stylet into host tissues, they become immobile. In the final generation of the growing season, settled first instar BWAs overwinter and begin the following year’s population. Cold winter temperatures (<−20 °C) can significantly reduce the overwintering population; however, refuge under snowpack insulates BWAs at the lower portions of the bole from lethal temperatures, reducing overwinter mortality [
10]. As BWAs progress through three instars toward adulthood they produce the protective wax-like “wool”, which surrounds their body and is the first sign of infestation. In the beginning stages of infestation, BWA and wool can be difficult to identify, even in detailed, field-based inventories. In addition, another woolly adelgid in this region,
Pineus abietinus Underwood & Balch, also infests true firs and appears identical to BWA [
17,
18]. Identifying the species must be carried out via a morphological examination of a slide-mounted specimen or via DNA extraction [
19]. As the BWA population establishes and grows, a symptom unique to BWA, gouting (abnormal swelling of branch nodes), can be used to identify and confirm this damage agent [
20,
21]. Once established in a stand, BWA populations may fluctuate year-to-year but will persist indefinitely so long as live hosts are available [
11]. Management strategies for BWA are limited to promoting non-host tree species via silvicultural treatment or planting. Attempts at slowing BWA population growth and damage due to insecticides are impractical at landscape scales. Silvicultural management options include increasing stand vigor and the removal of highly infested trees; however, evidence of their effectiveness is lacking.
Remote sensing is a critical tool for mapping the spatially and temporally explicit extent and severity of insect-induced disturbances in forest ecosystems [
22,
23,
24,
25]. By exploiting changes in spectral data over time (magnitude, timing, spatial patterns, etc.), we can identify areas where significant changes to forest canopy conditions have occurred. Ideally, these changes can be linked to one or more disturbance agents via a comparison to spatially coincident reference data, provided that a statistically robust relationship can be formed between the spectral change and the agent(s). This general approach has been demonstrated to be successful in a broad range of forest ecosystems [
26,
27,
28,
29,
30,
31]. However, the bulk of this research study has focused on bark beetles and canopy defoliators [
25]. Unlike these insects, which tend to produce relatively rapid and spatially clustered changes in forest canopy conditions and associated spectral responses, the complex spatiotemporal nature of BWA infestation has resulted in a relative paucity of remote-sensing-focused BWA studies. For example, tree health impacts from BWA infestation can unfold over the course of several years, and although mortality may be an eventual outcome, it is not guaranteed [
1]. Furthermore, the spatially diffuse nature of infestation attributed to wind-driven insect dispersal increases the likelihood that a diverse range of infestation severities may be present even within a single satellite image pixel of moderate spatial resolution (e.g., 30 m), obscuring the stand-level spectral signal of infestation [
16,
32]. Lastly, BWA is often only one among an array of agents acting in concert to damage a tree. For example, a tree can be weakened by BWA, but a subsequent infestation of bark beetles may be the dominant cause of mortality [
33]. Thus, attributing tree damage, both on the ground and from above, specifically and uniquely to BWA can be challenging.
Despite these limitations, a few studies have been successful at applying remote sensing to the study of BWA infestation. Franklin et al. [
34] demonstrated the novel capacity of classifying BWA infestation severity at the level of the individual tree; however, their approach relies on a relatively rare combination of high-spatial-resolution (0.5–1.0 m) and spectral resolution (288 bands) image data, limiting broad applicability. For example, satellite-based sensors with comparable spectral resolution, such as the DLR Earth Sensing Imaging Spectrometer (DESIS) and Hyperspectral Precursor of the Application Mission (PRISMA), have spatial resolutions of 30 m [
35,
36]. Cook et al. [
32] used reflectance spectra from a field spectrometer to characterize branch-level BWA infestation. However, they only characterized infestation in a binary fashion (infested vs. non-infested). From an ecological and management perspective, it would be more useful to be able to distinguish relative degrees of infestation (e.g., low, moderate, and severe) in order to understand relative impacts. Furthermore, while they did convolve field spectra to simulate satellite imagery, they note that using real remote sensing data would be challenging due to spectral mixing at the pixel level and the presence of multiple stressors on host trees obscuring the relatively subtle BWA-specific signal. To our knowledge, Hutten [
37] has been the only one carrying out an attempt at using a time series of remote sensing data to map BWA infestation at the stand level, finding that low-level change in the normalized burn ratio over time had statistically significant relationships with the presence of BWA infestation symptoms from field and aerial surveys. As with Cook et al. [
32], however, Hutten’s [
37] analysis did not quantify infestation severity on a continuous scale; instead, they merely distinguished between the presence and absence of BWA. Although these studies demonstrate promise, there remains a need to better understand the capacity for mapping BWA infestation on a continuous scale using widely available remote sensing time series data. Being able to map relative degrees of infestation provides land managers with a greater ability to prioritize mitigation efforts and understand potential future spread.
If remote sensing alone is insufficient for mapping BWA to a desirable degree of accuracy and precision, then perhaps additional geospatial data can augment the analysis. BWA damage has been shown to be strongly temperature- and terrain-dependent [
10,
37,
38,
39,
40,
41,
42]. Although many have noted this dependency, few have exploited it for mapping purposes. Hrinkevich et al. [
39] compared plot-level BWA infestation severity to a suite of climate variables derived from PRISM data [
43], finding summer and autumn temperature-related variables to be particularly important in predicting severity. This study provides evidence of the potential benefit of utilizing spatial data representing abiotic environmental factors, but the predictive power of their models was fairly low (R
2 = 0.24), and their results were generated at a relatively coarse spatial resolution (4 km). Thus, there remains a need to map infestation severity at a resolution and predictive accuracy that are of greater use to forest managers who require more spatially precise data to drive stand-level management decisions.
Remote sensing data can identify changes in forest conditions over time, although they often pose challenges with respect to distinguishing between causal agents with similar spectral signals [
26,
44]. Terrain and climate data can map susceptibility to insect infestation, particularly among species like BWA who have shown a strong climatic dependency, although susceptibility alone does not directly translate to certain infestations or their relative severity [
39]. Increasingly, there is a recognition that the combined use of remote sensing and spatially explicit representations of abiotic variables may exceed the capabilities of each data type used in isolation [
25,
28,
45]. Given BWA’s relatively subtle damage symptoms and demonstrated climatic dependency, this study aims to leverage the individual and combined use of spectral, terrain, and climate data to gain a robust understanding of the strengths and limitations of mapping BWA infestation severity.
The objectives of this study were to carry out the following:
Develop an accurate map of current BWA infestation severity for use by land managers, focusing on the relatively recent invasion of northern Utah;
Compare remote sensing-driven and terrain-/climate-driven approaches to mapping BWA infestation severity using a field-validated quantitative measure of stand-level severity;
Introduce a new approach for mapping BWA infestation severity that leverages individual strengths and overcomes the individual weaknesses of remote sensing and terrain/climate data;
Produce a quantitative accounting of landscape-level environmental drivers and identify key geospatial predictors of BWA infestation severity.
4. Discussion
In this study, we aimed to map the severity of an ongoing and relatively recent invasion of BWA into the subalpine fir forests of UWCNF in northern Utah. Given the limited research studies on mapping BWA infestation, we sought to test three different modeling approaches for deriving the most accurate, useful map for land managers. The spectral-only model, based solely on a time series of Landsat imagery, performed the worst of the three, although it was still able to account for approximately half of the variance in severity. We attribute the comparably poor performance to the fact that image data alone have a difficult time discerning between agents of change. Spectral reflectance and vegetation indices are excellent at identifying the extent and timing of disturbances and can certainly distinguish some types of disturbance from others (e.g., fire vs. harvesting) based on the differences in pre- and post-disturbance spectral characteristics [
44,
84,
85]. However, identifying the subtlety of BWA-induced damage, which occurs over several-year timescales, and distinguishing it from more pronounced but spectrally similar damage, such as the bark-beetle-induced mortality of subalpine fir or other codominant tree species, proved to be a challenging endeavor.
Figure 5 illustrates this challenge, providing examples of heavily infested forests that appear healthy and vice versa among our field plots. Nearly half of our plots exhibited damage that was dominantly attributable to non-BWA agents, further highlighting the complexity of teasing out causal factors in tree health decline in these forests. That said, our proportional agent scoring system, in concert with our approach for masking out disturbances such as fire, harvesting, and windthrow, enabled the mapping of BWA-specific infestation severity with some success, even using spectral data alone as the basis of predictions.
It is possible that the spatial and spectral resolutions of Landsat imagery may have been limiting factors to the performance of the spectral-only model. Within each 30 m pixel, even highly infested stands featured a range of canopy structures, mixtures of healthy and unhealthy trees, the presence of both host and non-host tree species, a diversity of understory vegetation, and variable ground surface materials, all of which can act to diminish the subtle spectral signal of infestation. The use of high-spatial-resolution imagery could reduce the amount of within-pixel mixing, potentially enabling the identification of tree-level infestation. Although successful evidence of this is limited in studies specific to BWA [
34], high-spatial-resolution image data from spaceborne and airborne platforms have proven effective when mapping the effects of other forest insects (e.g., [
86,
87,
88,
89]). Similarly, image data with higher spectral resolution than Landsat’s seven bands may have offered an opportunity to evaluate more spectrally distinct absorptive or reflective characteristics of BWA infestation. Once again, BWA-specific hyperspectral studies are limited [
32,
34], but studies of other insects have demonstrated promise, including the widespread and related hemlock woolly adelgid [
90].
The spectral-only model was also the most highly influenced by fine-scale spatial autocorrelation, as the model’s performance decreased notably with increasing cross-validation buffers. This is likely due to local patterns in vegetation composition over space since vegetation cover has a dominant effect on spectral reflectance in forest environments. As a result, a model built relative to one subalpine fir-dominated stand may not be directly applicable to another stand with a somewhat different vegetation assemblage. There is some debate in the spatial statistical literature as to whether or not this buffered leave-one-out cross-validation approach provides robust performance estimates. Wadoux et al. [
91] suggest that simple cross-validation (e.g., our 0 km buffer) tends to overestimate apparent model performance and buffered cross-validation tends to underestimate it. By presenting both cases, we feel that we have characterized our model’s performance fairly. One fundamental assumption in all spatial statistics is that close-proximity objects are more related to one another than distant objects, so by ensuring that training and test points have some distance between them, as in our buffered approach, we attempted to avoid the artificial inflation of apparent model performance [
76]. Despite the limitations of the spectral-only model, the variable importance and relationships observed among the predictor variables fall in line with expected trends in spectral response to BWA infestation and are corroborated by a body of previous literature [
32,
34,
37].
The terrain- and climate-only model had impressive predictive power, explaining approximately 75% of the variance in BWA infestation severity with an average prediction error of 0.078 based on only two temperature-based climate predictor variables. This suggests that BWA populations and subsequent damage to trees are heavily influenced by temperature. Areas with relatively high minimum summer temperatures (the most important climatic predictor of infestation severity) suggest a positive yet limited relationship between summer temperatures and BWA success. In fact, Greenbank [
10] found that the mean fecundity of BWA was the greatest from 8 to 24 °C in laboratory rearing studies, exhibiting increasing egg mortality at 26 °C and 100% egg mortality at 32 °C. Conversely, areas with colder temperature extremes and/or longer cold periods may limit both the inter- and intra-annual life cycle of BWA, thereby limiting infestation severity. This is illustrated by the importance of chilling degree days (the second most important predictor), which incorporates both the magnitude of difference in daily temperature from a baseline of 0 °C and the number of days below that same baseline. Thus, it is indicative of both extreme cold temperatures and/or extended cold periods. Greenbank [
10] found that overwintering first instar BWA mortality began at −20 °C, and no adelgids survived at temperatures below −34 °C (unless the adelgids resided below the snowline). Additionally, Greenbank [
10] concluded that colder climates may not support BWA infestations in the crown but that infestations could likely persist below the snowline and spread slowly. Note that although daily minimum temperatures may occasionally drop to lethal levels for BWA, the average minimum temperatures in
Figure 10 suggest that all subalpine fir forests of the UWCNF are at risk of BWA infestation. Hrinkevich et al. [
39] found very similar results to ours at a coarser resolution and broader spatial scale, with summer and autumn temperatures being the strongest predictors of BWA infestation. Although Mitchell and Buffam [
11] did not explicitly test climatic variables as the predictors of BWA infestation, their finding that lower-elevation sites, which we can infer were generally warmer, tended to have greater severity, which aligns with our results. Likewise, Hicke et al. [
92] found that warmer summer temperatures were associated with increased severity, mirroring our findings. One of the great benefits of our model is the fact that it can not only be applied to the fine-scale (30 m resolution) prediction of current infestation severity, as we carried out, but it can also potentially provide insight into future BWA infestation conditions. Future studies should aim to use the results we have generated to predict the potential areas of future BWA spread caused by climate change.
The combined model appeared to overcome limitations and enhance the individual strengths of the spectral-only and terrain- and climate-only models. The spectral model was good at capturing local-scale variability in forest damage (i.e., pixel-level changes in reflectance over time), but it was bad at capturing regional-scale trends with respect to BWA infestation severity due to poor change agent distinction. The terrain and climate model precisely performed the opposite, as it was driven by data downscaled to 30 m from a much coarser spatial resolution (800 m), which, itself, is a terrain-informed imputation of a sparse network of weather stations [
43]. For example, the Wasatch Mountains have certainly experienced a high degree of BWA infestation severity in recent years. The terrain and climate model accurately represents this regional phenomenon, but it depicts severity as a rather smooth gradient that is driven by trends in temperature. In reality, infestation can be highly variable at the local scale with adjacent forest stands featuring somewhat different levels of severity. This type of local variability would be missed by a purely climate-driven model and more likely to be captured by a spectrally driven model. Conversely, the Uinta Mountains have largely remained non-infested to date. The spectral-only map highlights several areas of low–moderate infestation, representing false positives that are likely caused by other damage agents presenting a similar spectral signal to BWA-induced damage, as highlighted in
Figure 5 and
Figure 12C. By building a model that incorporates spectral and climate data, forest degradation in BWA-prone areas is enhanced and that in BWA-resistant areas is diminished, producing a map that best captures both the local and regional patterns of BWA infestation severity. The prevailing relationships between the selected predictors of the combined model were very similar to those in the spectral and terrain/climate models. Spectrally, areas that featured increases in visible and shortwave infrared reflectance were associated with higher severity. Climatically, areas that were broadly characterized by warmer temperatures (e.g., higher minimum temperatures, lower subzero degree days, and more frost-free days) were associated with higher severity. To be sure, our combined model has not included all possible meaningful predictors of BWA infestation. It is certainly possible that the inclusion of predictors related to wind speed/direction (to understand BWA dispersal), forest canopy cover/basal area/biomass (to understand BWA forest structural preferences), and proximity to roads/development (to understand potential human-caused BWA spread), to name a few examples, could add predictive capacities to BWA mapping efforts in future studies.
All three models relied heavily on the use of a variable selection procedure known as VSURF [
71]. The goals of this process were as follows: (1) to increase model parsimony, which is generally good practice in statistical modeling; (2) to eliminate noisy or unhelpful predictors among a long list of candidates; and (3) to enhance the interpretability of model results according to prevailing trends in predictor–response relationships. While VSURF has been widely demonstrated to be effective toward these ends [
93], it is worth noting that the manner by which variables are eliminated could conceivably eliminate meaningful variables. For example, the third step of the three-step VSURF algorithm eliminates highly correlated variables. Inevitably, there was some degree of correlation among the many candidate predictors for each of our three models, particularly within the climate data (
Figure A5). While random forests are widely understood to be robust relative to multicollinearity [
94,
95,
96,
97,
98,
99], unlike parametric models such as multiple linear regression, VSURF may have removed variables that were highly correlated to, but slightly less important than, the final set of selected predictors in each model. Thus, our variable importance and selection results should be interpreted as follows: The variables that were selected were important, but the variables that were omitted were not necessarily unimportant. Moreover, with respect to our interpretation of the relationships between predictors and BWA infestation severity, our use of accumulated local effects was specifically aimed at addressing potential multicollinearity in the final predictor set for each model [
77]. Once again, while multicollinearity can negatively affect the ability to interpret the meaning of model coefficients in a parametric statistical context, assessing accumulated local effects of non-parametric random forests provides robust insight into predictor–response relationships.
In recognition of the fact that 58 plots are a relatively small sample in comparison to the large area over which we were carrying out the mapping (all subalpine fir in UWCNF) and that our plot protocol is merely one approach of many that can potentially quantify BWA infestation severity, we compared our results to an independent aerial survey dataset. The prevailing trends in severity measures and spatial distributions were well aligned between our maps and the areas identified by the USDA Forest Service ADS program as having been infested by BWA.
Figure 12A illustrates that ADS polygons with higher percent affected severity classes were generally mapped as exhibiting higher severity in our combined model map, although there is clearly a fair amount of spread within each class. It is impossible to determine if this is due to uncertainty in our maps or uncertainty in ADS data, but as we acknowledged in
Section 2.7, the spatial and thematic uncertainties inherent in ADS data can limit their use in serving as direct reference data for remote sensing analyses [
79,
80,
100]. Furthermore, the primary focal metric of damage in ADS data is mortality, a signal that is easily visually detected as red, orange, or brown conifer needles, which can be readily identified even from high altitudes. Although BWA infestation can kill trees, mortality was only abundant in the most severely infested stands in our field database. The driving indicators of severe infestation were gouting, crown deformities, and the presence of wool on the tree bole, all of which are much more difficult, or impossible, to identify from the air, as exemplified in
Figure 5. Thus, particularly low-level BWA infestations are likely poorly represented in ADS data.
Among the ADS damage polygons in subalpine fir forests of UWCNF, BWA was the most frequently attributed causal agent. The second-most common cause of damage was subalpine fir decline (SFD), which is also referred to as the subalpine fir mortality complex. As the name suggests, SFD is not a singular agent so much as it is a confluence of relatively poorly understood agents, including climatic factors, pathogens, and insects, all acting in concert to produce widespread subalpine fir mortality at times [
101,
102]. In our field data, it was nearly always the case that trees heavily impacted by BWA also featured other damage agents—especially bark beetles. Our
BDS vs.
ODS scoring system was designed to tease out BWA-specific damage, but the unexplained variance in each of our models (especially the spectral model) can likely be attributed to SFD. Additionally, given the apparent relationship between warmer temperatures and BWA in our terrain and climate and combined models, it is certainly possible that drought may have acted as an additional source of model confusion or even predisposed trees to BWA-induced damage. Northern Utah experienced extended drought conditions for several years preceding our field campaign, and this may have played a role in weakening trees with high moisture requirements, such as subalpine fir.
From a management perspective, it is often useful to describe severity categorically (e.g., “low”, “moderate”, and “high”). This type of categorization formed the basis of many of our field measurements of severity. Yet, the nature of our analytical approach yielded a measure of severity on a continuous scale. This scale ranges theoretically from zero, which would indicate not a single BWA-affected subalpine fir tree within a plot, to one, which would indicate the highest levels for all infestation metrics for every subalpine fir tree within a plot and no presence of other damage agents whatsoever. The plot-level quantitative measures of severity in our database ranged from 0 to 0.49. One might consider 0.49 to be moderate (since it falls roughly halfway between the theoretical minimum and maximum values), but in fact, this represents a heavily impacted site that one would certainly call “severe” (high mortality, severe gouting, etc.). Given that there are no hard definitions of what defines different categories of BWA infestation severity at the stand level, we have only presented our results on a quantitative scale. End users of the maps (e.g., forest managers, ecologists, and entomologists) may choose to apply their reclassification schemes to define thresholds that equate to categories of severity that suit their own needs. Of particular interest to forest managers may be the identification of low levels of infestation, as these areas may be targeted for management priority and neighboring areas may be highly susceptible to future infestation [
16].
There are a few limitations that warrant further discussion. First, there is inherent subjectivity in some of the infestation metrics that we measured. For example, distinguishing between light and moderate gout severity was sometimes a difficult judgment call. To address subjectivity, the same calls were carried out by the same person for all plots, although some variability is still certainly inherent in the data. Second, there is an innate challenge to evaluating some of those same metrics. For example, gouting is most easily observed on the ends of twigs with live foliage (
Figure 2). In stands where the live crowns of taller trees were not reachable from the ground level, identifying gout was difficult, particularly if gouting was light. Third, and perhaps most importantly, the results of this analysis should be treated as primarily relevant to the UWCNF, and extrapolation outside of the study area’s boundary should be done cautiously. One major reason is that we have not accounted for BWA spread over time. For example, BWA has been endemic in subalpine fir forests located north and west of the UWCNF for decades or longer. Conversely, BWA has not yet been identified in areas located south and east of UWCNF. However, a purely or mostly climate-driven model makes the assumption that BWA has had an equal opportunity to inhabit everywhere, irrespective of the actual time it takes for insect populations to spread over space. Accordingly, applying our models to north and west areas with comparable climatic conditions would likely result in the underestimation of severity, and applying our models to climatically comparable areas located south and east would likely result in the overestimation of severity. This points to a unique advantage of the inclusion of spectral data in the modeling process, as areas that have experienced vegetation change are the focus, rather than theoretical habitat suitability. In summary, we have presented a local model, both spatially and temporally, that is capable of predicting BWA infestation severity within the extent of UWCNF at present. We have not presented a global model that is capable of mapping BWA infestation over broader regions over longer timescales. Future research should aim to explore the best mapping practices for capturing broader-range variability in BWA infestation, perhaps incorporating some constraints that represent insect spread over time.