Classifying forest inventory data into species-based forest community types at broad extents: exploring tradeoffs among supervised and unsupervised approaches

Forest inventory data

Forest plot observations from across the conterminous U.S. were extracted from the FIA database (FIADB version 6.1; O’Connell et al. 2016). The FIA program uses a nationally-consistent, sample-based statistical design to quantify forest conditions across the U.S., and is the primary source for information about the status and trends of U.S. forest resources (Smith 2002). The FIA program samples all forest and other land uses, with one permanent plot established for every 2428 ha of land (Bechtold and Patterson 2005). FIA plots consist of four 7.2 m fixed radius subplots (0.067 ha or 67 m² each), with three subplots spaced 36.6 m apart in a triangular arrangement and one subplot in the center (O'Connell et al. 2016). Data collected for forested plots by field crews include the diameter and species of every tree stem in each plot. To protect sensitive plot information, especially on privately owned lands, the publicly available FIA database contains plot location information that has been altered slightly from the true location. We used actual plot locations to extract environmental variables at plot locations (described below), but show altered locations in Fig. 1. To ensure that we included the full set of plots for each state in the conterminous U.S., we selected the set of plots for each state that was used by FIA to produce a recent evaluation of forest conditions. For most states, the data represented evaluations from 2013, but for some states, evaluations were from 2012, and the evaluation for one state (Tennessee) was from 2009.

Because FIA plots have fixed points and layout, they can straddle, for example, multiple forest types, and to track those differences, multiple “conditions” can be identified within a plot. Each condition within a plot represents all or a portion of one or more subplots, and can have a different forest type and forest type group assigned to it (O'Connell et al. 2016). Therefore, to ensure that our sample units were each labeled with only a single forest type group and ensure our sample units represent a relatively homogenous tree species assemblage, we used the conditions within plots as our unit for classification. The FIA database contains information on each condition and the proportion of the plot it represents. For each plot, we used the single condition that made up the greatest proportion of the plot. In the case of ties, we picked one condition at random. We refer to these conditions as “plots” throughout this paper. We excluded plots labeled as being of planted origin. We also excluded plots labeled with FIA forest type groups “Exotic Hardwoods” or “Exotic Softwoods” because the empirical classification was based on native U.S. species only. We also excluded plots labeled as “Nonstocked” because those typically have few, if any, trees and we are specifically interested here in tree species composition.

For assigning plots to empirical classes, calculating species composition similarity, and analyzing indicator and dominant species, we needed information on species composition within each plot. For each plot, we assembled data on relative importance value by species for each plot we included. A species’ relative importance value is the average of its relative abundance and relative basal area compared to other species in the plot. Trees greater than 2.54 cm (1 in.) diameter at breast height were used to calculate importance value. The most rare species (those occurring on less than 250 plots) were excluded to avoid biasing the analysis toward extremely rare species (McCune and Grace 2002), and nonnative species were excluded because those species may respond to different environmental drivers than native species. We used the same data set of relative importance by tree species for the comparisons among classifications as well as indicator and dominant species analysis described below (see Costanza et al. 2017 for a list of all common and scientific names of tree species used here).

The three classifications and their assignment to FIA plots

The FIA class assignment for each plot came directly from the FIA database. Personnel from the USDA Forest Service assign forest types and forest type groups to its FIA plots using a decision-tree approach based on the relative stocking values of tree species in the plot, which are primarily a function of basal area (Arner et al. 2003). The combination of related species that comprises the highest stocking values is used to assign most of the forest types, and related forest types are aggregated into forest type groups. The FIA forest types are based on a number of sources, including a published set of types and groups that emphasize the cover of the dominant species and are largely derived from a classification by Eyre (1980). Thus, the FIA classification is a supervised approach, and the FIA groupings are static and well-recognized. The FIA forest types and forest type groups are widely used in maps of forest resources, ecosystem services, pest impacts and risk, and to inform initial conditions in forest landscape dynamics modeling studies (Ruefenacht et al. 2008; DeSantis et al. 2013; Duveneck et al. 2015; Healey et al. 2016).

The USNVC is a federal vegetation classification standard (FGDC 2008) that was developed independently of the FIA program, and applies to all vegetation. In the USNVC, forest types are classified at multiple levels using both tree and non-tree species, growth forms and ecology (the EcoVeg approach of Faber-Langendoen et al. 2014). Types have been recognized based on a combination of local to regional empirical data, literature, and expert judgement (Franklin et al. 2015; USNVC 2016). USNVC types have largely been formally described and published (USNVC 2016; http://www.usnvc.org), but the classification is dynamic and open to peer-reviewed changes. The classification has used been to understand the distribution of vegetation types locally to nationally (Hoagland 2000; Matthews et al. 2011; Belote et al. 2017), and by land management agencies for conservation planning (Franklin et al. 2015). Although the USNVC standard is national, the classification is based on vegetation plot data from a variety of sources, along with ancillary environmental characteristics such as soil wetness and disturbance information (Franklin et al. 2015).

Recently, a supervised decision tree, or key, was developed to assign FIA plots to one level of the hierarchy, the macrogroup level, in the eastern U.S. (Menard et al. 2017). The macrogroup is a mid-scale unit (5th level, approximately at the scale of FIA forest type groups) and is based on sets of diagnostic and dominant plant species and growth forms that reflect biogeographic differences in composition and sub-continental to regional differences in ecological factors (FGDC 2008). The macrogroup classes were developed based on ancillary local data and expert opinion for all vegetation types, not just forests, but the supervised key applies them to tree data in the FIA database. Expansion of the key to the entire country has been proposed but not yet implemented. Because the bulk of FIA data only contain tree species data and limited ecological data, the key is necessarily restricted to those data, though the ecoregion within which each plot is located is also used as a proxy for non-tree species and ecological factors. The USNVC classes were assigned to FIA plots based on the relative importance value information for each species described above, including trees greater than 2.54 cm (1 in.) diameter at breast height within the plot. Although the key assigns FIA plots to both “cultural” (plantations and other planted stands) and “natural” (non-planted) types, only plots falling into one of the natural macrogroups were included here.

The third classification of FIA plots was a recently developed set of empirically-derived forest community types (Costanza et al. 2017). The empirical classification used an unsupervised hierarchical method to cluster tree species composition information within FIA inventory plots across the conterminous U.S. (Costanza et al. 2017). The clusters were based on the relative importance value for species in each plot, using the same set of species described above and trees greater than 2.54 cm (1 in.) diameter at breast height (Costanza et al. 2017).

Thus, unlike the FIA classes, which rely most strongly on one, two or three dominant species, the empirical classes are defined based on the composition of a larger group of species in a plot. This classification was developed to facilitate the study of forest dynamics over time in response to global change scenarios (Costanza et al. 2017).

The empirical classification is hierarchical, and any level of the hierarchy with any number of clusters can be chosen to fit the objectives of a given study. We used the level with 29 clusters spanning the conterminous U.S. because that was one level that we previously identified as optimal based on standard diagnostic metrics for hierarchical clusters (Costanza et al. 2017). From our previous work, we had a set of cluster “seeds” that had been assigned to one of the 29 empirical clusters (Costanza et al. 2017). We used the relative importance values from those seeds as the training set in k-nearest neighbor classification algorithm to assign the current set of FIA plots to those 29 clusters. In other words, we assigned each FIA plot in the current study to one of the 29 clusters based on the similarity of its relative importance values to those in the data set that was used to develop the empirical clusters.

The FIA and empirical classifications apply to all FIA plots in the conterminous U.S.; however, the key for USNVC macrogroups has only been developed for FIA plots in the eastern U.S. Therefore, we assigned all three of the alternative forest community type labels for each plot in the eastern U.S., and two labels for plots outside the eastern U.S. (Fig. 1). A set of 70,425 forested plots in the eastern U.S. and a total of 117,813 forested plots spanning the conterminous U.S. were used.

Correspondence among classifications

We compared the classifications to determine the degree of correspondence among classes. We completed four total two-way comparisons: three in the eastern U.S., among the USNVC, the FIA, and the empirical classifications, and one across the entire U.S., between the FIA and empirical classifications. In each case, we determined the relative similarities among classes in the pairwise comparison, both in terms of the frequency distributions of plots in the classes and in terms of the average species composition of classes. We first constructed two-way cross-tabulated contingency tables showing the frequency distributions of plots in the classifications, and input those tables into correspondence analysis. Correspondence analysis is similar to principal components analysis but is used for analyzing tables of count data (Greenacre 2013). It provides a graphical representation of the relationship between two categorical variables via an examination of how classes load onto important correspondence axes. Specifically, it can point out the classes that are most strongly associated with one another.

For each class in each classification, we also calculated the average species composition in terms of relative species importance values across all FIA plots that were assigned to the class; that is, we identified the centroid of each class in multivariate species space. Importantly, although each classification scheme used slightly different information from the FIA database to label plots (for example, the FIA classes used relative stocking, etc.), the same relative importance value by species as described above was used to compare the classes here. For classes in each classification, the centroids of relative importance values were used to construct matrices of pairwise similarities between classes, using 1 - Bray-Curtis dissimilarity. We examined those similarities and visualized them using heat maps. For all pairwise comparisons, we omitted classes from each classification that had fewer than ten plots in them to simplify comparisons because initial tests of correspondence analysis indicated that those small classes had disproportionately large effects on the results.

Indicator and dominant species

Because the classification techniques differed in the extents to which they relied on dominant or characteristic tree species, we examined how well each classification could be characterized by indicator and dominant species. We conducted indicator species analysis and calculated species dominance values for all classes (Dufrene and Legendre 1997; Frieswyk et al. 2007). We used the same set of tree species as described above, with nonnative and the most rare species removed. Indicator species analysis finds both dominant and non-dominant species that have high specificity and fidelity to a given class using a permutation test to assign an indicator value and associated p-value to those species (Dufrene and Legendre 1997) and any set of alternative classifications can be compared based on those metrics (Dufrene and Legendre 1997). Classifications ranked highly by indicator species analysis will have at least one significant indicator species associated with each class, higher total numbers of indicator species overall, larger sums of all significant indicator values, and lower p-values on average. We used those metrics in our comparison of the classifications here. We standardized all indicator species metrics that were based on sums or counts by dividing by the number of classes in each classification.

We also calculated a species dominance index (SDI) for every species in each class, following Frieswyk et al. (2007). Dominant species are those that comprise a large proportion of the species assemblage in a plot, and thus the dominant species in a species assemblage are often responsible for much of the assemblages’ ecological functioning and ecosystem services (Frieswyk et al. 2007; Hildebrand et al. 2008; Le Roux et al. 2014). The SDI ranges from 0 to 1, with 1 being a perfectly dominant species within a given class. A species with a high SDI value for a given class will tend to: (1) have high canopy cover on average across all plots within the class, (2) occur with few or no other species in plots within the class, and/or (3) only occur in a small number of plots within the class but tend to have high cover where it occurs (Frieswyk et al. 2007). The SDI is typically based on canopy cover, but here we use relative importance values (described above) in place of cover. We averaged the SDI values for the top most dominant species in every class within the classifications to determine which classifications tended to produce classes with more dominant species.

Predicting the classifications using environmental variables

We used the nonparametric random forest algorithm (Breiman 2001; Cutler et al. 2007) to determine how well each classification scheme could be predicted using environmental variables alone, and to examine the most important environmental predictors for each classification scheme. Each of the classification schemes was used as the response variable in a separate random forest model. The random forest algorithm fits many classification trees to a data set using a subset of predictors and a bootstrap sample of the data, then combines the results (Prasad et al. 2006; Cutler et al. 2007). In the random forest algorithm, the data not used to construct each classification tree, called the out-of-bag observations, are used to determine the error rate of each tree. Variable importance is based on the difference between each tree’s error rate and a permutation of the tree without the given variable (Cutler et al. 2007, 2012).

As predictor variables for random forest models, we compiled spatial data on local and landscape soil characteristics, recent historical climate, landscape, and topography from ancillary sources (see Additional file 1 for a list of all variables used). We overlaid FIA plots on these spatial data layers to extract the values for plot locations. We focused on variables that have been shown to be important in other recent studies of tree distributions, vulnerability, and climate change (e.g., Iverson et al. 2008; Rogers et al. 2017). Soil variables were taken from the 10-m resolution Gridded Soil Survey Geographic Database (gSSURGO; Soil Survey Staff 2017a), where available. We used variables for the dominant component in the topmost horizon in gSSURGO data. For a few counties in the eastern U.S. and several places in the western U.S., gSSURGO spatial data were not available, and we substituted data from the polygon-based State Survey Geographic Dataset (STATSGO2; Soil Survey Staff 2017b). We also included landscape Productivity Index and Drainage Index as measures of landscape-scale soil properties of each site (Schaetzl et al. 2009, 2012). We used average recent historical climate (1979–2015) data from the Multivariate Adapted Constructed Analogs (MACA) v2 METDATA data set (Abatzoglou 2013) to derive 19 bioclimatic variables (Hijmans et al. 2005), as well as annual averages of potential evapotranspiration and solar radiation. We also used the MACA data to calculate growing degree days (Sork et al. 2010), and annual summed moisture index (Koch and Coulston 2015). A set of seven topographic variables were derived from the 30-m resolution National Elevation Dataset (Gesch et al. 2002). Three landscape condition metrics were also used as predictors. One was ecological landscape condition, which assesses the level of human stressors surrounding a plot (Hak and Comer 2017). The other two were forest area density metrics, which measure the area and contagion of forests in a 65.6-ha window surrounding each FIA plot based on the 2011 National Land Cover Database (Riitters and Wickham 2012; Homer et al. 2015).

We ran random forest models for the FIA and empirical classifications across the U.S., as well as all three classifications in the east. We set the number of trees in each random forest model to 2000 and used all other defaults in the randomForest package in R (Liaw and Wiener 2002). We also used the default output from randomForest which assigns a predicted class to every data point based on the majority of votes across all trees in the forest. The random forest algorithm aims to minimize the overall error rate across all classes; thus, we report the average out-of-bag error rates. We also report Cohen’s Kappa statistic of interrater reliability here for each classification (Cohen 1960).

All data manipulation and analysis was done using R 3.4.0 (R Core Team 2017) and the following contributed packages: ca (Nenadic and Greenacre 2007), class (Venables and Ripley 2002), factoextra (Kassambara and Mundt 2017), irr (Gamer et al. 2012), labdsv (Roberts 2016), randomForest (Liaw and Wiener 2002), RColorBrewer (Neuwirth 2014), raster (Hijmans 2016), and tidyverse (Wickham 2017), and vegan (Oksanen et al. 2016).

Correspondence among classifications

We completed four two-way comparisons of plot frequencies and species composition: three in the eastern U.S. and one nationwide. We only included classes with greater than ten plots for this portion of the analysis, which resulted in eliminating 11 plots from the national data set and 21 plots from the eastern U.S. data set. In each of the four comparisons, some classes had relatively good analogs in terms of species composition and cross-tabulated frequencies of plots and some did not. Here, we point out some of the classes with high and low correspondence in each case using three summary methods: biplots from the first three dimensions of each correspondence analysis (Figs. 2 and 3), heat maps of species similarity (Fig. 4), and the full contingency tables of cross-tabulated plot frequencies (Additional file 3).

The comparison between the empirical and USNVC classifications in the eastern U.S. shows some broad similarities, but few one-to-one analogs in terms of plot frequencies. Biplots from correspondence analysis indicate a few cases in which one empirical class was associated with several USNVC classes (Fig. 2a). For example, the balsam fir-quaking aspen empirical class was associated with five USNVC classes. In this case, all the 1082 plots in the North American Boreal Conifer Poor Swamp USNVC class were classified in the balsam fir-quaking aspen empirical class, but that empirical class also included plots in several other USNVC classes (Additional file 3). These differences likely occurred because the same tree species exist in both upland and wetland boreal and sub-boreal forests. The slash pine-longleaf pine empirical class was associated with the Pond-Cypress Basin Swamp and Longleaf Pine Woodland USNVC classes in terms of plot frequencies, again crossing upland and wetland USNVC classes. The loblolly pine-sweetgum empirical class was associated with the Southern & South-Central Oak-Pine Forest & Woodland, the Southeastern North American Ruderal Flooded & Swamp Forest, and the Southern Mesic Mixed Broadleaf Forest USNVC classes, thus crossing, upland, wetland and ruderal forest classes. In terms of average species composition, the highest similarity among pairs of classes in the empirical and USNVC classifications was between the slash pine-longleaf pine empirical class and the Longleaf Pine Woodland USNVC class (Fig. 4a). Despite having low correspondence in terms of plot frequencies, the loblolly pine-sweetgum empirical and Southern Mesic Mixed Broadleaf Forest USNVC classes were relatively similar in species composition (Fig. 4a). The sugar maple-red maple empirical class had a relatively high average similarity and was associated in terms of plot frequencies with two USNVC classes: the Appalachian-Interior-Northeastern Mesic Forest and the Appalachian-Northeastern Oak-Hardwood-Pine Forest & Woodland (Figs. 2a and 4a). Several empirical and USNVC classes had low similarities in species composition to all classes in the other classification. Those included the honey mesquite-Pinchot juniper and chokecherry-Pacific dogwood empirical classes, and the North American Boreal Conifer Poor Swamp USNVC class, likely because those empirical classes were largely determined by western species. Interestingly, in contrast to the high correspondence in terms of plot frequencies, the similarity in species composition between the balsam fir-quaking aspen empirical and the North American Boreal Conifer Poor Swamp USNVC class was low.

The comparison between the FIA and USNVC classifications in the eastern U.S. also shows some broad similarities, but few one-to-one analogs in terms of plot frequencies. Biplots from correspondence analysis show a few cases in which classes from the two classifications were associated on a one-to-many or many-to-many basis (Fig. 2b). The oak/pine and oak/hickory FIA classes were associated with two USNVC classes: the Southern & South-Central Oak-Pine Forest & Woodland, and the Appalachian-Northeastern Oak-Hardwood-Pine Forest & Woodland. Here the USNVC separates two classes based on overall floristic and biogeographic differences, whereas the two FIA classes emphasize the mixed conifer-deciduous versus pure deciduous overstories. The aspen/birch and spruce/fir FIA classes were also both associated with two USNVC classes: the Laurentian-Acadian-North Atlantic Coastal Flooded & Swamp Forest, and the Laurentian-Acadian Mesic Hardwood-Conifer Forest, thus crossing upland and wetland USNVC classes, but distinguishing deciduous from conifer. There were two other strong associations between one FIA class and three USNVC classes. The only one-to-one association in terms of plot frequencies was between the longleaf/slash pine FIA class and the Longleaf Pine Woodland USNVC class (Fig. 2b, Additional file 3). Average species similarities among classes showed the highest pairwise similarity was between the Longleaf/slash pine FIA class and the Longleaf Pine Woodland USNVC class (Fig. 4b). The Maple/beech/birch FIA class had relatively high similarity on average with two USNVC classes: Laurentian-Acadian Mesic Hardwood-Conifer Forest, and Appalachian-Interior-Northeastern Mesic Forest. Other high average similarities in species composition occurred between the Southern Coastal Plain Floodplain Forest USNVC and Oak/gum/cypress FIA classes, and between the Laurentian-Acadian Pine-Hardwood Forest & Woodland USNVC and White/red/jack pine FIA classes. Two FIA classes had low similarities with all USNVC classes, namely the Pinyon/juniper and woodland hardwoods classes. These classes were largely based on western tree species. Other FIA classes had moderate similarity with several USNVC classes. For example, the loblolly/shortleaf pine group was moderately similar in species composition to four USNVC classes including Southern Mesic Mixed Broadleaf Forest. All the USNVC classes were at least moderately similar to a minimum of one FIA class.

In the eastern U.S., the comparison between the empirical and FIA classifications shows both broad similarity and one-to-one analogs for some classes. The honey mesquite-Pinchot juniper empirical class was associated with the woodland hardwoods FIA class, indicating that those two classes were good analogs in terms of plot frequencies (Fig. 3a, Additional file 3). The balsam fir-quaking aspen empirical class was associated with three FIA classes – the spruce/fir, aspen/birch and white/red/jack pine – suggesting that the empirical class is more heterogeneous than the FIA class. The slash pine-longleaf pine empirical class was associated with the longleaf/slash pine FIA class. When comparing the FIA and empirical classes in the eastern U.S. in terms of species composition, similar analogs stand out. Three pairs of classes had the highest similarities: (1) the balsam fir-quaking aspen empirical and aspen/birch FIA classes; (2) the slash pine-longleaf pine empirical and longleaf/slash pine FIA classes; and (3) the honey mesquite-Pinchot juniper empirical and woodland hardwoods FIA classes (Fig. 4c). Six empirical classes and five FIA classes in the eastern U.S. had low similarity in species composition on average with all classes in the other classification. Examples of those classes are the black willow and cedar elm empirical classes, and the other eastern softwoods and other hardwoods FIA classes. The largest empirical classes in terms of numbers of plots had relatively high similarities in species composition with more than one FIA class. For example, the loblolly pine-sweetgum empirical class was most similar in species composition to the oak/pine FIA class, but it also had relatively high similarities with other FIA classes, including the loblolly/shortleaf pine and oak/gum/cypress FIA groups.

When the empirical and FIA classifications were compared for FIA plots across the U.S., there were necessarily some similarities to the eastern U.S. comparison because eastern U.S. plots were included in the full set of U.S. plots. However, there were also several important differences, and we focus on those differences here. There were only two sets of classes that were associated with one another in terms of the first three dimensions of correspondence analysis, and all included classes that did not occur in the east (Fig. 3b). The Utah juniper-two needle pinyon and alligator juniper-Arizona white oak empirical classes were associated with the pinyon/juniper FIA class, and three empirical groups containing western oak species were associated with the western oak FIA class. In terms of average species composition, there were many pairs of classes with low to moderate similarity, and only a few cases of relatively high similarity (Fig. 4d). The highest similarities overall were those described above for the eastern U.S., but the Utah juniper-two needle pinyon empirical class also had high similarity in average species composition to the pinyon/juniper FIA group.

Indicator and dominant species

Predicting the classifications using environmental variables

The average out-of-bag error rates and Kappa statistics for all classifications show that the set of environmental variables included here best predicted the empirical classification, either when considering plots across the conterminous U.S. or the subset of eastern U.S. plots (Table 3; full confusion matrices are in Additional file 5). The empirical classification across the U.S. had the lowest error rate and highest Kappa value, and the empirical classification applied to eastern plots had the second lowest and highest values, respectively. Metrics for the FIA classes across the U.S. were third best overall, but metrics for the USNVC were better than those of the FIA classes for eastern plots.

A somewhat similar set of variables tended to be among the most important for predicting all classifications (Fig. 5). The mean temperature of the coldest month, slope, and the proportion of forest in the landscape were the top three most important for predicting the empirical and FIA classifications, whether for all plots or just eastern plots (Fig. 5a-b, d-e). Mean temperature of the coldest month, precipitation of the warmest quarter, and precipitation seasonality were the most important variables in predicting the USNVC classification (Fig. 5c). Soil and other climate variables tended to be next most important in all cases. Landscape soil variables and aspect-related topographic variables tended to be least important for predicting all classifications. For the empirical and FIA classifications, the order of variable importance changed slightly when considering all plots versus eastern U.S. plots, but the most and least important variables stayed relatively constant.