Seven clusters or groups were chosen to describe the community as this reflected the ecological separation of stands and an agglomerative stage at which a greater increase in the distance coefficient was observed (Fig. S1, see the Supplementary material). The classification efficiency was 52% at agglomeration stage 40. Monte Carlo tests for the non-metric multidimensional scaling ordination revealed a better than random solution (P < 0·05) for each of one to six dimensions. The final stress value for a two-dimensional solution was 11·0% (0·00001 = final instability, 158 = no. of iterations). All seven communities classified using clustering techniques clearly separated on the first two ordination axes (Fig. 2).
Figure 2. Non-metric multidimensional scaling ordination of 47 stands. Axes reflect ordination of standardized species data. Letters indicate community membership determined from agglomerative clustering. Abbreviations are as follows: A, open muskeg; B, shrub-rich treed muskeg; C, mature black spruce; D, black spruce; E, deciduous; F, mixed forest; G, mixed forest–black spruce.
Download figure to PowerPoint
The communities differed in the abundance of arboreal lichen (F = 9·135, n = 47, P < 0·001) and reindeer lichen (F = 21·831, n = 47, P < 0·001) available to woodland caribou. Multiple comparison tests revealed that the black spruce community (class D) contained a significantly greater abundance of arboreal lichen than communities characterized as open muskeg (class A), shrub-rich treed muskeg (class B) and deciduous (class E) (Fig. 2). The class D community was dominated by black spruce and contained minimal amounts of balsam fir, white birch and Populus spp. in the canopy. Stands younger than 100 years in age occurred as this community type and the canopy ranged from sparse to dense. The mature black spruce community (class C) contained a significantly greater percentage cover of reindeer lichen than all other communities. This community was represented by mature black spruce stands, more than 100 years of age, with canopy closure less than 67% and an abundance of arboreal lichens. No differences in lichen abundance were detected among stands of deciduous (class E), mixed forest (class F) and mixed forest dominated by black spruce (class G). Class E stands were dominated by Populus spp. and balsam fir and classes F and G were dominated by black spruce, Populus spp. and balsam fir, with lesser amounts of white spruce Picea glauca ((Moench) Voss), white birch, Populus spp. and jack pine Pinus banksiana (Lamb.). Detailed descriptions of all communities can be found in Appendix S1 (see the Supplementary material).
canonical ordination and variance decomposition
Three structural variables and five species variables were retained, following forward selection, in the FRI data set: height, stocking, age, Betula papyrifera, Abies balsamea, Populus spp., Picea mariana and Picea glauca. Although not statistically significant, the balsam fir and poplar species categories were included because of the importance of these species in the preceding community classification. Three structural variables and four habitat class variables were retained, following forward selection, in the Landsat data set: fractal dimension, spectral bands 3 and 4, shrubs-hardwood, shrubs-disturbed, shrub-rich treed fen and open bog.
The variation explained by information from FRI and Landsat was high relative to the variation in community species composition identified using unconstrained CA. The total variation in the community species composition was 2·445 (sum of all unconstrained eigenvalues). The first two axes of CA accounted for 39·3% of this variation. Separate canonical ordinations of the FRI and Landsat factors, irrespective of any covariance, revealed that both accounted for at least 80% of the observed variation in the community vegetation (Table 3). For both explanatory sets the correspondence between CA and CCA was high for the first axis but declined by up to 10% for the second axis (Table 3).
Table 3. Variation (%) in vegetation data explained by the first two axes of correspondence analysis (CA) and canonical correspondence analysis (CCA) (n = 47)
|1 FRI||2 Landsat|
|CCA (%CA)||20·5 (82)||21·5 (86)|
|CCA (%CA)||10·8 (75)||10·9 (76)|
|CCA (%CA)||31·3 (80)||32·4 (82)|
The first CCA axis derived using FRI explanatory variables showed a strong contrast between stands dominated by poplar (class E), stands dominated by black spruce or mixed conifer (classes C, D and G) and stands of open and treed muskeg (classes A and B). Poplar (−0·68), height (−0·59) and stocking (−0·62) showed the greatest correlation with axis 1, while black spruce (−0·69) was the most highly correlated explanatory variable on axis 2 (Fig. 3). Stands separated along axis 1 according to their height/stocking character, with open bogs and sparse-canopy black spruce stands (classes A, B, and C) falling on the right side of the axis. Mixed forest black spruce stands (classes D and G) occurred in the middle range, while stands characterized by greater height and canopy closure (class E) fell on the left side of the axis. On axis 2, stands separated according to the percentage of black spruce in the canopy, with open bog and deciduous stands occurring near the top (classes A, B, E), followed by mixed forest stands (class F) and then black spruce-dominated stands (class D) clustered near the bottom.
Figure 3. Canonical correspondence ordination of forest resource inventory (FRI) explanatory variables (vectors) (n = 47). Letters indicate community membership determined from agglomerative clustering. Axes reflect ordination of log-transformed species data. Abbreviations are as follows: BW, Betula papyrifera; BF, Abies balsamea; HT, height, PO, Populus spp.; SB, Picea mariana; STKG, stocking; SW, Picea glauca; A, open muskeg; B, shrub-rich treed muskeg; C, mature black spruce; D, black spruce; E, deciduous; F, mixed forest; G, mixed forest–black spruce. Only significant independent variables (P = 0·05) are shown, based on forward selection using canoco.
Download figure to PowerPoint
Ordination of stands constrained by Landsat explanatory variables showed similar separation along each axis to the FRI ordination (Fig. 4). Spectral band 3 (0·77) and the shrubs, hardwood forest class (−0·51) had the greatest correlation with the first axis, while band 4 (0·67) and fractal dimension (−0·52) had the strongest correlation with axis 2. A clear separation was apparent between deciduous forest, muskeg and conifer stands. Pooling of stands revealed that muskeg (classes A and B) had a significantly lower fractal dimension than deciduous (classes E and F) or conifer stands (classes C, D and G) (F = 22·29, n= 47, P < 0·001). Band 3 shared a similar vector orientation to the Landsat open bog variable. Spectral scores for band 3 were significantly greater for muskeg (classes A and B) than deciduous (classes E and F) or conifer stands (classes C, D and G) (F = 22·29, n= 47, P < 0·001). Conifer habitat classes in the Landsat variable set did not explain a significant amount of variation in stand species composition. However, the vector for band 4 was orientated similar to the SB vector in the FRI ordination (Fig. 4). Black spruce stands (class D) had significantly lower band 4 scores than mixed forest (class F); and mixed conifer–black spruce forest (class G) had significantly lower values than mixed forest (class F) or deciduous forest (class E) (F = 8·457, n= 47, P < 0·001).
Figure 4. Canonical correspondence ordination of Landsat explanatory variables (vectors) (n = 47). Letters indicate community membership determined from agglomerative clustering. Axes reflect ordination of log-transformed species data. Abbreviations are as follows: A, open muskeg; B, shrub-rich treed muskeg; C, mature black spruce; D, black spruce; E, deciduous; F, mixed forest; G, mixed forest–black spruce; Band 3, green TM spectral score; band 4, red TM spectral score; FD, fractal dimension; SD, shrubs-disturbed; SH, shrubs-hardwood; STF, shrub-rich treed fen. Only significant independent variables (P = 0·05) are shown, based on forward selection using canoco.
Download figure to PowerPoint
Variance decomposition of stands revealed that approximately 45% of the total variation in community species composition (CA) was explained by stand (FRI) and landscape-level (Landsat) factors (Fig. 5). In stage 1, stand-level factors (31·3%) showed similar effects to landscape-level factors (32·4%) when each was considered separately. In addition, both showed similar independent effects (FRI = 12·2%, Landsat = 13·2%). The two-way overlap of both data sets revealed that 61·0% of the explanatory power of stand-level factors was confounded with landscape factors, and 59·1% of the explanatory power of landscape factors was confounded with stand-level factors.
Figure 5. Partitioning of the explanatory power of forest resource inventory (FRI) and Landsat (LS) factors on community species composition in the Clay Belt region, Ontario, Canada. The area of each cell is proportional to the variance accounted for by that component. Total inertia = 2·445. Numbers indicate the percentage of total species variation (total inertia) accounted for by each component CCA analysis. All CCA analyses were significant (P < 0·05) based on Monte Carlo permutations.
Download figure to PowerPoint
In stage 2 of the decomposition, total explanatory power of the FRI species and structural factors (39·1%) was similar to the total explanatory power of the Landsat habitat class and structural factors (40·3%). Stand-level species composition in the FRI data set had greater explanatory power than structural attributes, independently accounting for almost twice as much of the species variation. Twenty-eight per cent of the explanatory power of FRI species factors was confounded with structural factors, while 40·7% of the explanatory power of structural factors was confounded with species factors. Conversely, Landsat structural components explained a greater proportion of variation than habitat factors, accounting for almost half of the total explained variation of all components (Fig. 5). Notably, fractal dimension was significantly less for FRI stands than Landsat stands (t = −85·079, n= 46, P < 0·001).