Forest Edge Regrowth Typologies in Southern Sweden—Relationship to Environmental Characteristics and Implications for Management

Södra stambanan (Southern main line, hereafter SML) is the main railroad track in southern Sweden. It connects Malmö (55.609814N; 13.00115E) the third largest city, with Stockholm (59.330497N; 18.05603E), the capital. Applying a 1000 m buffer around SML in analysis of national GIS data (© Lantmäteriet i2014/00764, © Naturvårdsverket, i2014/764; © SMHI i2014/00764) shows that the altitude of the railroad ranges from 0 to 336 m above sea level (mean ± standard deviation (SD) 105.15 ± 78.30 m). Mean annual precipitation ranges from approximately 600 to 900 mm and the cumulative temperature sum per year ranges between 1272 and 1789 °C (mean ± SD 1494.88 ± 131.78 °C). This gives an overall slightly humid climate. The positive product of mean precipitation, after subtraction of evapotranspiration during the vegetation period, ranges between 33 and 156 mm (mean ± SD 91.71 ± 34.15 mm), with an overall decreasing trend toward the east. The soil type varies widely and includes peat soils, glacial tills, and sedimentary soils ranging from clayish to coarse sandy and gravel types. The variations in soil and macroclimate are reflected in differing landscape composition along the railroad, ranging from plains with intensive agriculture and urbanization (as little as 6% forest cover), through mosaics of small-scale agriculture and forest patches, to forest landscapes (as much as 94% forest cover). The dominant land cover is forest, with a total cover of 45% divided between conifer (26%), deciduous (8%), mixed (4%), and clearcut/regenerating forest (7%).

After severe problems with wind-felled trees and related disruptions to railroad traffic, the SML management corridor was one of the first in Sweden to be expanded from 10 to 20 m on both sides of the track, by clearing forest edge vegetation in the NMC.

A field inventory of SML was conducted in autumn 2011, four growing seasons after clearance, at 78 random sites located on the forest edge as defined by the National Inventory of Landscapes in Sweden (Gallegos Torell 2011) (Fig. 1). At each site, a plot of 35 × 20 m was established so that one half covered the NMC and the other half covered the bordering forest stand (hereafter denoted STAND), resulting in a NMC and STAND zone each measuring 10 × 35 m (Fig. 2). The border between STAND and NMC was set as the average outer line of tree trunks, in accordance with the NILS field protocol (Gallegos Torell 2011).

Within the plot, four transects were established 10 m apart, perpendicular to the railroad (Fig. 2). Along each transect, five circular subplots (1–5) were set so that the first two plots were located in the NMC (the first had its center 9 m outside the forest stand border, the second its center 4.5 m outside the stand border), the third was located with its center at the border between the STAND and NMC zones, and subplots 4 and 5 were located in the forest STAND (4.5 and 9 m inside the forest stand, respectively), resulting in a total of 20 subplots per site (Fig. 2). In each subplot, woody vegetation was sampled at two levels: (a) within a 0.5 m radius, species and number of woody plants <1 m in height were counted; (b) within a 1 m radius, species, height, and number of woody plants ≥1 to ≤5 m in height were recorded. Furthermore, in (b) the mean height of the field layer (10 cm intervals) was measured and field layer cover (10% intervals) was estimated. For the STAND zone, a third level (c) was added where within a radius of 2 m, circumference at breast height (CBH) was measured for all specimens >5 m in height.

Based on the field data, two woody species matrices were developed at plot level. (i) A species matrix for the NMC was established using subplots 1–3 from each transect (i.e., a total of 12 subplots) with species represented as counts of stems (sampling level b), but with specimens <1 m (sampling level a) restricted to one count per species and subplot to downgrade the importance of small seedlings with little probability of full survival for all counts. (ii) A species matrix for the STAND was also established, where species were represented by the CBH for subplots 3–5 (sampling level c).

Using the species data derived from the field work described above, Wiström and Nielsen (2016) applied ordinations to identify ten environmental variables at different scales that are significant for woody regrowth in the NMC and the bordering forest STAND. These variables represent aspects of soil conditions, climate, vegetation, and landscape structures as summarized below and in Table 1.

In relation to the objectives of the study, the following analytical steps were applied in sequence:

All multivariate analysis was performed in R (R Core Team 2014) using the additional packages vegan (Oksanen et al. 2013), cluster (Maechler et al. 2014), labdsv (Roberts 2013), mvpart (Therneau et al. 2013), and MVPARTwrap (Ouellette and Legendre 2013), with syntax and analytical approaches adapted from Borcard et al. (2011). Species matrices for NMC and STAND were analyzed separately. The different analyses are described in detail below.

For both species matrices, four of the most common hierarchical clustering approaches were applied using Chord-transformed data (which removes the total abundance per site, i.e., the response of the species to the total productivity of the sites) and Euclidian distances: (i) Single linkage agglomerative clustering, which is good for detecting gradients; (ii) complete linkage agglomerative clustering, which is efficient in finding distinct groups; (iii) average agglomerative clustering using the unweighted pair group method with arithmetic mean (UPGMA), which can be seen as intermediate to the above methods; and (iv) Ward’s minimum variance clustering, which defines groups so that the within-group sum of squares is minimized, often giving tightly bound clusters, although these may not always reflect the underlying data set (El-Hamdouchi and Willett 1989; Borcard et al. 2011).

To compare the different clusterings, the cophenetic correlation coefficient between the original dissimilarity matrix and the cophenetic matrix from the clustering was calculated. They were also plotted together in Shepard-like diagrams (Legendre and Legendre 1998) with a lowess smoothing function. Based on this, the most appropriate clustering was selected for further analysis.

To assess the optimal number of clusters according to silhouette widths, the Rousseeuw quality index was used, together with fusion plots where number of clusters was plotted against node height. Using the selected number of clusters, the amount of misclassification in silhouette widths was assessed through silhouette plotting (Rousseeuw 1987).

To visualize the species and their abundance in relation to the clustering and dendrogram, the tabasco function with “log” as scale range in vegan, which displays compact community tables with an interface to the heatmap function, was applied. This gave a color image of the species data abundance ordered in relation to the clustering.

The output of the clustering and tabasco heatmap was then examined for tree species dominating the different clusters for the NMC and STAND zones. Six tree species clearly dominated in the different clusters. An “idealized management data set” was then created where these six dominant tree species with a high probability of driving regrowth toward an abrupt edge profile and homogeneous structure were removed from the NMC data set (as if they had been cut by selective thinning). Acknowledging that the six species will regenerate generatively and/or vegetatively, this idealized data set enabled analyses of the regrowth in relation to a management scenario. The clustering was then re-run to determine and assess the relative abundance of remaining species, i.e., the idealized management data set enabled assessment of the presence and abundance of smaller tree and shrub species in the different clusters that, if promoted by management, could form the “building blocks” for development of graded forest edges. Furthermore, the idealized management data set enabled the identification of other minority tree species that have to be addressed in long-term management. When sites had to be filled with a pseudospecies to enable the analysis to be run, this indicated a lack of other species than the six dominant tree species.

Multivariate regression trees (hereafter shortened to “regression trees”) were used to relate the regrowth clusters to the environmental variables and identify the variables that drive the woody regrowth into different species compositions (De’ath and Fabricius 2000; De’ath 2002). Significant species for the division of groups and the final grouping were identified by combining the regression trees with a Chord pre-transformation and the species indicator values developed by Dufrene and Legendre (1997).

Species are the building blocks of the forest edges. However, it is not only the species composition, but also the overall structure of the edges and its relation to regrowth types, that are of interest. One important aspect is to compare differences in vegetation structure between the regression tree groups, in order to enable in-depth analysis without directly including species assembly. Therefore, some key responses relating to the forest edge structure and edge effect were modeled using the placement of the five different subplots perpendicular to the forest edge. This was done using mixed general linear models in SAS 9.3 Proc Mixed procedure. The structural aspects, i.e., responses, for the different models were number of stems, maximum height of stems, number of seedlings (i.e., height <1 m), average field layer coverage, and average field layer height. These were calculated as the mean value of the transects for each subplot level at the sites. Square root transformation was used on the response when needed to fulfil the assumptions of the model, which were checked through inspection of diagnostic plots of the conditional residuals. For each response, a model was fitted with the regression tree classification (class variable with five levels) and subplot (class variable with five levels), and their interactions as explanatory variables. Regression tree classification was nested under “Site” to avoid pseudoreplication. “Site” was modeled as a random variable. To investigate and handle the probable dependence between subplots, each model was compared concerning model fit with Akaike information criterion (AIC) using either a standard variance components structure (VC) or a first-order autoregressive structure (AR1) for the covariance structure for random effects. To analyze interaction effects, the SLICE statement was used in a partitioned analysis of the least squares means, which is described as an analysis of simple effects by Winer (1971). As such, the SLICE statement looks at the predicted values across subplot for different levels of classification, and vice versa.