*3.2. Modeling of Formations*

Ten different methods of grouping formations were tested (Table 3). In methods 1 and 2, all sample plots are used and each formation is modeled individually, and in method 2, the DEM is excluded from environmental variables. In methods 3–6, formations E, F and G (oak, linden, broadleaf forests) are combined, in method 3, the DEM is excluded from the environmental variables and in method 4, only spectral reflectances of July and September are left. Method 5 combines the H and J formations (birch and gray alder forests). Methods 7–9 apply spatial rarefication of occurrence data. In method 7, only the formation points A (spruce) are rarefied. In methods 8 and 9, different rarefication distances are applied for formations A, B, C, D and H. Two pairs of ecologically similar formations are merged: E with F and J with K. In method 9, the DEM is excluded from environmental variables.


**Table 3.** Modeling results for the 9 methods.

<sup>1</sup> Method 8 including Palsar-2 dataset. <sup>2</sup> Broadleaf in Forest Plan includes maple, ash and elm. <sup>3</sup> Birch including gray alder.

The results of the eighth method of grouping of formations are closest to the ratio of tree species given in the Forest Plan (FP) [17]. The modeling is re-run with settings of the eighth method and addition of the Palsar-2 dataset (Figure 1). The results of modeling the dominant formations are in good agreement with the data of the Forest Plan: birch 30.1% (39.6% in the FP), spruce 25.1% (24.4% FP), pine 18.5% (20.7% FP), aspen 5.1% (8.4%) and oak with broad-leaved (E + F) 2.9% (1.78% FP). The proportions of linden (G) and alder (J + K) are overestimated by 7.3 and 3.1 times respectively, compared to the data of the Forest Plan (Figure 2). Several possible reasons for this discrepancy are suggested. First, there is a lack of field data. Naturally, in the absence of preliminary stratification, such formations will rarely be encountered during field routes. At the same time, despite the fact that the number of points of the rare formations has been conditionally brought to the level of 80–100, they nevertheless still have a significant role of autocorrelation: often these points are located in spatial clusters, which reduces the quality of models. However, another factor is also important. In the course of field work, it was repeatedly noticed that the forests indicated on the forest inventory maps as birch are in fact alder forests.

**Figure 2.** Cartographic model of forest formations and non-forest habitats.

According to the confusion matrix (Table 4), the classification accuracy is 0.46. The best classification quality is for alder forests. This is noteworthy in the context of the above difference with the official data of the Forest Plan. A low level of matching with field data is found for the formations of aspen (I) 0.31, spruce forests (A) 0.33 and oak broad-leaved forests (E + F) 0.34. Aspen forests (I) are poorly separated from birch (H) forests. For spruce forests (A), matching problems are associated with a closely related formation: spruce-small-leaved forests (B). Oak broad-leaved forests (E + F) are poorly separated from spruce-small-leaved (B) and birch (H) forests. The use of the Palsar-2 dataset allows to increase overall accuracy from 0.44 to 0.45, and it gives a rather significant accuracy increase for pine (D), 0.57 to 0.62, and linden (H), 0.35 to 0.45 (confusion matrix without the radar dataset is not demonstrated here).


**Table 4.** Confusion matrix model of forest formations.

#### *3.3. Modeling of Association Groups*

Despite the uneven spatial distribution of points, modeling of association groups is performed without spatial rarefication. Table 5 shows the results of modeling of association groups. The average level of points matching between all association groups is 0.29 and it ranges from 0.03 to 0.69. On average, a low percentage of points matching is typical for coniferous groups of associations (A–D) and broad-leaved conifers (F). Oak-spruce (E), linden (G), birch (H) and aspen (I) have an average level of recognition. The best quality is for gray alder (J) and black alder (K) formations.

According to the assessment of the spatial distribution of the identified association groups of forest communities (Table 5), the largest area (12.5%) is occupied by communities of derivative birch and aspen forests (Formations H and I) with a predominance of the mesotrophic and hydromorphic series (Bh, MhBh, Gm). The proportion (4.7%) of birch (H) grass-marsh forests (Gm), which are distinguished by strong recreational disturbance, is also high. Within the pine formation (D), the maximum proportion (about 5%) is occupied by communities (DShG and Sh), which tend to succession towards boreal spruce forests. The same pattern is observed in the formation of mixed spruce-small-leaved forests (B), occupying an area (4.72%) where their successional dynamics are also directed towards the restoration of spruce communities of groups (DShG and Sh). Another part of these communities of formation B (3.3%), in terms of the composition of the ground layer (ShBh and Bh), has a tendency to succession towards broad-leaved communities with a nemoral composition of the ground layer. It is obvious that this group is of artificial origin. The area of communities within the spruce forests (A) is small and varies in the range of 0.47–1.37%.

Table 6 shows the proportion of the area occupied by non-forest habitats.



*Forests* **2020**, *11*, 1088 Italic: level of points matching.


**Table 6.** Percentage of total area by non-forest land cover types.
