2.1 Revised EUNIS habitat classification

A list of the individual habitats belonging to these six groups is provided in Table 1, and habitat factsheets with their descriptions and corresponding phytosociological alliances of EuroVegChecklist (Mucina et al., 2016) are in Appendix S1. We prepared the lists of corresponding alliances based on expert judgement by comparing the basic characteristics of the EUNIS habitats with the EuroVegChecklist alliances. These lists can help understand the content of individual EUNIS habitats to those scientists and practitioners who are familiar with European phytosociological classification. However, although the EUNIS habitat classification was, to a large extent, inspired by the phytosociological classification system, it developed independently of it, which implies that the phytosociological alliances are often not nested within habitat types. The “one-to-many” or “many-to-one” relationships of the EUNIS habitats to the EuroVegChecklist alliances are much more common than simple one-to-one matches.

Within the six revised habitat groups, we selected only those habitats that could be defined based on floristic criteria (Table 1, Appendix S1). We did not develop formal definitions for those habitat types that represent mosaics of several different habitats (e.g., wooded pastures) because their complete structure cannot be represented by single vegetation plots. We also did not consider those forest habitats that are defined based on the management practice or successional stage but do not differ floristically from related types with different management or in different successional stages (e.g., T41 Early-stage natural and semi-natural forest and regrowth or T43 Coppice and early-stage plantations). We were able to define plantations of non-site-native trees (those planted at sites where they would not occur naturally) but unable to define plantations of site-native trees because their floristic composition is usually indiscernible from that of natural forests.

2.2 Data sources

The primary source for producing characteristic species combinations and maps for EUNIS habitats was a data set of European vegetation-plot records (henceforth “vegetation plots” or “plots”). Such plots typically contain a full list of vascular plant species, often also a list of bryophytes and lichens, estimates of cover abundance of each species and various additional sources of information on vegetation structure, location and environmental features in the plot (Dengler et al., 2011). These plots were extracted from the EVA database (Chytrý et al., 2016; accessed on 19 May 2020), and several other databases not included in EVA (see the full list of databases used in this study in Appendix S2). The geographical scope was the whole of Europe (including the European part of Russia), Azores, Madeira, Canary Islands, Anatolia, Cyprus, Georgia, Armenia and Azerbaijan. The data set contained plots representing both the target habitat groups (coastal, wetlands, grasslands, shrublands, forests and man-made) and non-target habitat groups (marine, inland surface water and inland sparsely vegetated habitats). The latter groups are not in the focus of this study because the revision of their classification is not yet complete or published; still, plots sampled in these types were needed to assure correct classification of the habitats that are transitional between the target and non-target habitat groups. We excluded the plots that reported only species composition without cover-abundance information for individual species. Further, we excluded plots smaller than 1 m², larger than 1,000 m², without geographical coordinates and those with reported uncertainty of the coordinates larger than 10 km. The plots with a missing indication of location uncertainty were retained, assuming most of them were within 10 km from the indicated coordinates. The resulting data set contained a total of 1,261,373 georeferenced plots. The data set was prepared using the Turboveg 3 program (Hennekens, 2015) and analysed using the Juice 7.1 program (Tichý, 2002). It is the most extensive data set of vegetation plots ever analysed (compare Bruelheide et al., 2019).

The taxon names in this data set originated from various national and thematic international databases, most of them managed in Turboveg 2 (Hennekens and Schaminée, 2001), which use different taxon lists with partly inconsistent taxon concepts and names. Taxonomy and nomenclature were unified using Turboveg 3 in two steps. Firstly, the names from the original databases were interpreted by regional botanists, considering the taxonomic concepts and nomenclature used in the focal region of each database. This step was important because it solved regional differences in the use and meaning of some taxon names. In this step, taxon lists of most of the European vegetation-plot databases were matched to accepted names in the SynBioSys Taxon Database, an unpublished working database of taxon names and concepts used in the EVA project (Chytrý et al., 2016). Secondly, the names of vascular plants from the SynBioSys Taxon Database and the names from the original databases that did not match any name in the SynBioSys Taxon Database were translated to the nomenclature of the Euro+Med PlantBase (Euro+Med, 2006–2020; ww2.bgbm.org/EuroPlusMed), using a complete list of accepted names and synonyms of European and Mediterranean vascular plant taxa provided by the Berlin-Dahlem Botanical Garden and Botanical Museum in February 2020. The names of bryophytes and lichens followed the SynBioSys Taxon Database.

The cover of individual species was, in most vegetation plots, recorded using a cover-abundance scale (70% of plots were recorded using a variant of the Braun-Blanquet scale; Westhoff and van der Maarel, 1973). We transformed all of these scales to the arithmetic mid-point percent cover values corresponding to the individual cover-abundance classes following the default conversion of the Turboveg 2 program (Hennekens and Schaminée, 2001). For the seven-grade Braun-Blanquet scale the transformation was r + 1 2 3 4 5 → 1% 2% 3% 13% 38% 63% 88%, for the nine-grade Braun-Blanquet scale r + 1 2m 2a 2b 3 4 5→1% 2% 3% 4% 8% 18% 38% 63% 88%, and for the Domin scale + 1 2 3 4 5 6 7 8 9 10→1% 2% 3% 4% 13% 23% 29% 42% 63% 88% 99%. The occurrences of the same species in different layers of the same plot were merged, and their percentage covers recorded in different layers were combined using an algorithm that assumes random overlap of covers or different species, resulting in cover values that cannot exceed 100% (Chytrý et al., 2005; Jennings et al., 2009; Fischer, 2015; further referred to as the “Jennings‒Fischer formula”). As a result, each species was present only once in the data set. This step was necessary because information on layers was not recorded in many plots.

2.3 EUNIS-ESy: An expert system for identifying EUNIS habitats in vegetation-plot databases

The new classification expert system EUNIS-ESy (= EUNIS Expert System) was developed for identifying coastal, wetland, grassland, shrubland, forest and vegetated man-made habitats of the EUNIS Habitat Classification based on species composition and cover abundances of particular species or species groups. In the habitats that are difficult to distinguish based on purely floristic criteria, plot-location criteria were added.

EUNIS-ESy is based on formal definitions of habitats written as logical formulas in an editable script stored as a TXT file (Appendix S3). The computer program that runs the expert system evaluates all the plots of a vegetation database and checks for each of them whether it meets the conditions of one or more of the formal definitions of habitats included in this script. If a plot matches a definition of one habitat, it is assigned to this habitat. In an ideal case, habitats should be mutually exclusive, and each plot should be assigned to one and only one habitat. However, in reality, some plant communities have a transitional composition corresponding to two or even more habitats. The plots representing such communities are simultaneously assigned to all of these habitats. Other plant communities with idiosyncratic or impoverished species composition may be unable to be assigned to any habitat and remain unclassified by the expert system. Nevertheless, the expert system was prepared with the aim of allowing a large majority of plots to be assigned unequivocally to a single habitat.

The expert system script of EUNIS-ESy (Appendix S3) is divided into three sections, which represent successive steps in the analysis. Section 1 merges selected taxon names. In most cases, it merges subspecies, varieties or forms to the species level. We did so in order to improve the consistency of taxonomic concepts across the data set because some authors recorded only species while others also recorded infraspecific taxa. Further, we merged taxonomically difficult groups of species containing many misidentifications into species aggregates.

Section 2 enumerates species belonging to individual species groups that characterize particular habitats or groups of habitats and are used in the formal definitions of habitats. A single species can be assigned to more than one group. The initial lists of species included in the groups were compiled from relevant phytosociological literature, national habitat handbooks, personal field experience, data from our previous EUNIS reports (Schaminée et al., 2013, 2014, 2016a, 2016b), factsheets of the European Red List of Habitats (https://forum.eionet.europa.eu/european-red-list-habitats/library/terrestrial-habitats; see Janssen et al., 2016) and other sources. These initial lists were critically revised and extensively modified based on multiple classification trials with successive versions of EUNIS-ESy (Section 2.4).

Section 3 comprises definitions of habitats written as logical formulas that combine taxonomic specifiers, relational operators and threshold abundance criteria, which can be joined as required by the logical operators AND, OR or NOT. Technical description of the syntax of the expert system is provided by Tichý et al. (2019: their Appendix S1).

2.3.3 The hierarchical structure of the expert system

The expert system was developed hierarchically. Each habitat definition was assigned a priority degree. When the expert system is running, the definitions with the highest priority are applied to the data set first, and the plots that meet the requirements of these definitions are assigned to the habitats, while other plots remain unclassified. Then, the definitions with lower priority are applied to the remaining unclassified plots.

The current study deals only with the EUNIS habitat groups that were revised so far, i.e. N (Coastal), Q (Wetlands), R (Grasslands), S (Shrublands), T (Forests) and V (Man-made). However, the hierarchy of the expert system has been designed to include the vegetated habitats from the other groups (A — Marine, C — Inland surface waters, and H — Inland sparsely vegetated) once their revision is finished and published. Preliminary definitions of these habitats were developed and included in the expert system, which is important for separation of the target habitat groups from the non-target groups. The codes and concepts of these habitats in the expert system correspond to those used in the European Red List of Habitats (Janssen et al., 2016). However, these preliminary definitions were not tested and therefore not included in the results of the current study.

In some cases, we created two definitions with different priority levels to define a single habitat. The narrower definition is applied at a higher priority level. It is usually based on the occurrence or a high total cover of species from a functional group that comprises species narrowly specialized to the habitat. This definition classifies the plots that are typical examples of the particular habitat, but it leaves many less typical plots of this habitat unclassified. Subsequently, a broader definition is applied at a lower priority level to the unclassified plots. This definition is based on a discriminating species group and classifies the plots that are less typical examples of the habitat but still possess more features of this habitat than of any other habitat. Such a two-step approach is needed for habitats in which the occurrence of narrowly specialized species is a sufficient criterion for habitat assignment even if such species have a low cover. If only definitions based on discriminating species groups were used, some of the plots of these habitats could be misclassified.

For example, the habitat R11 Pannonian and Pontic sandy steppe is defined by two formulas. First, a narrower definition is applied at a higher hierarchical level:

• <#TC R11-Pannonian-and-Pontic-sandy-steppe-specialists GR 15> NOT <#TC Trees|#TC Shrubs GR 15>

This means that the total cover of the functional species group R11-Pannonian-and-Pontic-sandy-steppe-specialists, including a selection of narrow ecological specialists of this habitat, should have a cover greater than 15%, and the total cover of trees and shrubs should not be greater than 15%. Then, a broader definition of the same habitat is applied to the plots that were not yet classified by any formula with a higher priority level:

• (<##Q +04 R11-Pannonian-and-Pontic-sandy-steppe> AND <#03 +04 R11-Pannonian-and-Pontic-sandy-steppe>) NOT <#TC Trees|#TC Shrubs GR 15>

This means that the sum of square-rooted percentage covers of a discriminating species group of the Pannonian and Pontic sandy steppe (including both the narrow specialists and frequently occurring less-specialized species) should be greater than the sum of square-rooted percentage covers of the discriminating species groups of any other habitat, and the plot should contain at least three species of this group, and the total cover of trees and shrubs should not exceed 15%.

The expert system contains habitat definitions assigned to eight priority levels (Figure 1). The definitions at the highest priority level (8) are applied first in the classification, while the definitions at the lowest priority level (1) are applied last:

• Level 8: Coastal habitats dominated by woody plants (i.e. coastal heaths, dune scrub and dune forests; habitats N18 to N1G) and Macaronesian heath (S43) are defined based on a high cover of dominant species or a high total cover of functional groups of dominant species (dwarf shrubs, shrubs and trees) in combination with the occurrence on the coast, on coastal dunes, or in Macaronesia, respectively.
• Level 7: Other (i.e. herbaceous) coastal habitats (N11–N17 and N1H–N1J) and marine habitats of the tidal zone dominated by vascular plants (A25a–A25d) are defined based on the discriminating species groups of habitats within these two groups, in combination with occurrence on the coast or in coastal dunes.
• Level 5–6: The habitat group H (inland sparsely vegetated habitats) is defined based on a cover not greater than 30% in combination with the occurrence of at least one species specialized to the habitats of this group. Individual habitats within this group are provisionally defined based on their preliminary discriminating species groups at level 5. In the future, level 6 with narrower definitions of these habitats based on specialist species will be added.
• Level 4: Some coastal herbaceous (group N), some grassland (group R), all shrubland except the Macaronesian heath (group S), all forest (group T) and some man-made (group V) habitats are classified using definitions based on a high cover of characteristic dominant species or functional groups of characteristic dominant species, or the presence of a specified minimum number of species narrowly specialized to individual habitats.
• Level 3: Habitat groups of shrublands (group S) and forests (group T) are defined based on the dominance of shrubs, dwarf shrubs or trees. Vegetation plots that have not been previously classified to specific habitats within these groups are classified directly to these broad groups.
• Level 2: All the non-shrubland and non-forest habitats that were not defined before are defined based on the discriminating species groups.
• Level 1: Habitat groups of wetland (Q, separated into the groups of Qa — mires and Qb — helophyte beds), grassland (R), man-made (V), inland surface water (C) and inland sparsely vegetated (H) habitats, without assignment to any specific habitat, are defined based on discriminating species groups. Vegetation plots that have not been previously classified to specific habitats within these groups are classified directly to these habitat groups.

2.4 Iterative evaluation and optimization of the expert system

The expert system and formal definitions of individual habitats therein were created based on expert opinion combined with iterative improvement, which used information from the evaluation of the results of successive classification trials. A preliminary version of the expert system contained the initial species groups (Section 2.3) and the first version of formal definitions for a subset of habitats belonging to the same habitat group (e.g., Forests). The formal definitions were proposed by the authors of this paper, who considered the content of each target habitat and the options provided by the formal language of the expert system. The aim was to propose such definitions that would include most plots belonging to the habitat and none or very few plots not belonging to the habitat. This preliminary version of the expert system was applied to a data set of European vegetation plots in the Juice program. The resulting classification was evaluated by the experts, focusing on false positive (a plot does not belong to the habitat but is assigned to it) and false negative (a plot belongs to the habitat but is not assigned to it) classification results. The classification could only be validated based on the judgement of human experts because there is no standard of correct plot-level classification. Therefore the plots that the expert system assigned to individual habitats were checked by several experts from different countries, who were specialists in different habitats. These plots were also mapped, and the experts paid special attention to geographically outlying plots and to the absence of plots in areas where the habitat was expected. If the experts identified misclassified plots, the expert system was modified to avoid such misclassifications. The modifications were made either to the content of the species groups or to the assignment rules in the formal definitions. For species groups, the species that contributed to misidentifications were identified and removed from the groups, while other species that might contribute to the correct habitat identification were added to the relevant groups. For formal definitions, the structure of the formulas or thresholds used in the formulas were changed. This process was repeated many times until misclassification identifiable by the experts were eliminated. Once the final classification was achieved for one habitat group, the first versions of formal definitions of another habitat group were added, and the iterative optimization process was repeated.

2.5 Characteristic species combination

In phytosociology, “characteristic species combination” is defined as a combination of diagnostic species and species with higher constancy that together define a vegetation unit (Braun-Blanquet, 1964). Here we use this term as an umbrella for the three types of species that Chytrý and Tichý (2003) introduced to characterize vegetation types: diagnostic, constant and dominant species. Diagnostic species (Whittaker, 1962; Westhoff and van der Maarel, 1973) are species with occurrences concentrated in a particular habitat, being absent or rare in other habitats. As such, they are useful as positive indicators of the habitat. However, diagnostic species may be absent from the habitat at many sites. Constant species are species that occur frequently but not necessarily exclusively in a particular habitat: some of them may be generalist species that are also frequent in other habitats. Dominant species are those that often reach high cover in a particular habitat, thus determining the habitat physiognomy.

For the purposes of computing the characteristic species combination for each EUNIS habitat at Level 3 of the classification hierarchy, we used a data set of all the plots classified at this level, including the vegetated marine (A), inland surface water (C) and inland sparsely vegetated (H) habitats, which were defined provisionally in the expert system. We performed a stratified resampling of this data set (Knollová et al., 2005) to balance the spatially uneven sampling effort across Europe, i.e. a high concentration of vegetation plots in relatively small areas contrasting with low density or absence of vegetation plots in other, often large areas. This procedure should reduce the bias in identification of diagnostic, constant and dominant species, especially for those species that are frequent in heavily sampled areas but rare or absent elsewhere. The stratification was applied to a data set of those vegetation plots that were classified by the expert system to Level 3 habitats, excluding the plots classified to Level 1 (habitat groups) but not to Level 3 habitats. All the plots in this data set were assigned to geographical grid cells of 5 min × 3 min of longitude × latitude (corresponding to approximately 6.0 km × 5.5 km at 50° N). If a cell contained more than one plot belonging to the same habitat, one randomly selected plot was retained, while the others were removed. If such resampling resulted in <20 plots per habitat across the whole data set, some of the previously removed plots were selected randomly and returned to the data set, ensuring that the total number of plots of the habitat was 20. Habitats with fewer than 20 plots in the whole data set were not resampled. The resampled data set contained 233,352 plots, i.e. 28% of all the plots classified to Level 3 habitats, but it was more balanced and more representative than the original data set.

Diagnostic species were determined based on species fidelity, i.e. the degree of concentration of species occurrences in each group of plots representing a Level 3 EUNIS habitat. Fidelity was calculated using the phi coefficient of association (Sokal and Rohlf, 1995; Chytrý et al., 2002) standardized as if each habitat was represented by the same number of plots (Tichý and Chytrý, 2006). The species with a value of phi greater than 0.15 for a particular habitat were considered as diagnostic for this habitat. This threshold was selected arbitrarily as a compromise between a stringent selection of few species with high diagnostic value (if phi was higher) and a lax selection of many species with weak diagnostic value (if phi was lower). However, the concentration of species occurrences in the habitat, even if expressed by a high value of the phi coefficient, may not be statistically significant for some habitats represented by a low number of plots in the data set. Therefore, the statistical significance of the species–habitat association was tested using Fisher's exact test (Sokal and Rohlf, 1995), and if not significant at p < 0.05, the species was excluded from the list of diagnostic species (Tichý and Chytrý, 2006).

Constant species were defined as those with a constancy (= percentage occurrence frequency) of at least 10% in the target habitat. This threshold is much lower than usually used for constant species of vegetation types in phytosociology. However, a lower value is needed for EUNIS habitats than for vegetation types, because many habitat types comprise several vegetation types occurring across broad geographic ranges with varying species composition; as a result, few species have a higher constancy across the whole habitat.

Dominant species were defined as those that occurred with a cover greater than 25% in at least 5% of vegetation plots classified to the target habitat. This means that a species is considered as dominant even if it does not belong to the tallest vegetation layer, and a single plot can have more than one dominant species. Conversely, a habitat can have no dominant species, especially if it has sparse vegetation cover.

Records of taxa identified only to the genus level and records of epiphytic lichen species were removed from the characteristic species combinations. Records of other non-vascular plants (bryophytes and non-epiphytic lichens) were retained because many of these species are important ecological indicators. However, as they were not recorded in all plots, their calculated constancy values were likely underestimated. Their fidelity can be either underestimated (if they were sampled only in some proportion of plots of the habitat) or overestimated (if bryophytes and lichens were more often sampled in some habitats than in others). A solution would be to compute constancy values only for plots where bryophytes and lichens were recorded (or would have been recorded if present). However, this was not possible because vegetation plots without records of bryophytes and lichens in most cases do not contain information whether these species were really absent or just not recorded. Therefore, we reported the values calculated for bryophytes and lichens based on all the plots, but we emphasize that these values can be inaccurate and have to be interpreted with caution.

As a quality test of our results, we made a formal comparison of characteristic species combinations computed for the EUNIS forest habitats with an earlier established list of indicator species for French forest habitats prepared on the basis of a different data set (Gégout et al., 2009). This exercise, performed by national experts (J.-C. Gégout and L. Maciejewski), revealed a high degree of correspondence between both lists, thereby indicating the reliability of the characteristic species combinations computed in our study.

2.6 Habitat distribution mapping

Distribution maps for individual habitats were prepared by plotting the location of all vegetation plots classified to individual habitats (before stratified resampling) on a map. All the maps were checked for outlying locations, which in most cases pointed out either an error in coordinates or misidentification of an important species that led to an erroneous classification to a different habitat. Errors were corrected, new classification prepared, and both the characteristic species combinations and distribution maps were updated.

Because of a strong geographic bias in the available European vegetation plots, especially their low density in northern and eastern Europe, we indicated the locations of plots belonging to individual habitats on grid maps showing the regional density of plots belonging to the particular habitat group. Such maps indicate whether the absence of occurrences of the habitat in a region is likely real or caused by the absence of data from the region (Figure 2).

2.7 Expert system software tools

A software tool to apply the expert system script to a data set of vegetation plots was developed within the Juice 7 program and, in a simpler form that does not contain all the functions, also in the Turboveg 3 program. The syntax of the expert system was described by Tichý et al. (2019: their Appendix S1). The code for applying the expert system in the R program (www.r-project.org) was developed by Bruelheide et al. (https://git.loe.auf.uni-rostock.de/misc/ESy).

4.1 EUNIS-ESy and the development of the EUNIS Habitat Classification

EUNIS-ESy is the first tool that automatically classifies vegetation plots across Europe to habitat types of the EUNIS Habitat Classification. The development of this expert system represents a major step forward in the applicability of the EUNIS Habitat Classification for nature conservation survey, planning, monitoring and reporting on the international, national and regional levels.

EUNIS comprises concepts of individual habitat types that resulted from discussions of international teams of experts and public consultations with national experts and practitioners, organized by the European Environment Agency (Rodwell et al., 2018). Therefore, the aim of the current study was not (and could not be) to revise this classification or concepts of individual habitats within it. Our aim was to develop formal definitions that would closely match the concepts of individual habitats and enable correct assignment of vegetation plots to these habitats. However, the work on this expert system was done in parallel with the EUNIS revision process in 2013–2019, and various experiences from developing formal definitions were fed back to the revision process and influenced its outcome (Schaminée et al., 2012, 2013, 2014, 2016a, 2016b, 2018, 2019, 2020). Further refinements of the formal definitions and expert system were made during the preparation of the current paper, based on the feedback from an international team of co-authors. Therefore, the results presented here are an update of the work that was previously summarized in the reports cited above.

The present paper deals with the six habitat groups for which the EUNIS classification has already been revised (Schaminée et al., 2018, 2019, 2020): coastal, wetland, grassland, shrubland, forest and man-made habitats. These habitat groups represent a large majority of the European terrestrial habitat types. Revisions have not yet been completed or published for the other groups, including marine habitats, inland surface waters, inland unvegetated or sparsely vegetated habitats, constructed, industrial and other artificial habitats, and habitat complexes. Many habitats of these remaining habitat groups are based entirely on the abiotic or non-plant features or are habitat complexes comprising several different plant communities. Therefore, the expert system approach based on floristically defined vegetation types cannot be used to identify them. Still, there are some habitats in these remaining habitat groups for which it will be possible to develop formal definitions and add them to the expert system, once these habitats are revised in the process guided by the European Environment Agency. In the marine habitats, the expert system would presumably also work with non-plant benthic species if data were available in a suitable form. These tasks remain for the future. Nevertheless, the current expert system includes preliminary definitions of the non-revised habitats from the remaining groups. In this way, its structure is prepared for the inclusion of new habitat definitions, which will replace the current preliminary definitions.

4.2 Comparison with other expert systems

The expert system EUNIS-ESy presented here is not the first one developed for the classification of European vegetation plots. Mucina et al. (2016) provided expert-based lists of diagnostic species for European vegetation classes defined in EuroVegChecklist and included them in an expert system. In this EuroVegChecklist expert system, diagnostic species were used directly as discriminating species in our terminology (see section 2 Methods), which provides an acceptable classification for many plots. However, unless diagnostic species lists are optimized for the purpose of identification of vegetation types, the classification error rate in expert systems is relatively high (Tichý et al., 2019), which is the case for the EuroVegChecklist expert system. Also, the EuroVegChecklist expert system does not consider cover or dominance of individual species or groups of species belonging to specific layers (e.g., trees or shrubs). Therefore it often fails to discriminate between vegetation types with similar species composition but different physiognomy (e.g., heathland vs forest with heath-like undergrowth).

In our work on EUNIS-ESy, we found that habitat/vegetation types defined by species of the dominant growth form or the growth form of the highest layer are most accurately defined by threshold covers of the dominant species or a dominant species group, in some cases in combination with other criteria. This holds true especially for shrubland and forest habitats, which are often defined by the dominance of a single tree or a small species group of trees or shrubs. In contrast, most types of grasslands, which often have a single layer of vascular plants with the dominant species changing from place to place, are better defined by discriminating species groups. Using different approaches to defining habitats within different habitat groups, EUNIS-ESy provides a more accurate (as assessed by expert judgement) classification of individual vegetation plots than the EuroVegChecklist expert system, which applies the same classification approach across all vegetation formations.

Another expert system on the European scale was developed by Giannetti et al. (2018) for the classification of European Forest Types produced by forestry experts as a tool for sustainable forest management (EEA, 2006; Barbati et al., 2014). This expert system classifies European forests to 14 types based on the dominant tree species, their basal area and information derived from plot location, e.g., altitude, biogeographical region or occurrence in wetland areas. Like EUNIS-ESy, it is a rule-based expert system (Grosan and Abraham, 2011) using the information provided by human experts, both on the dominant species and location, as classification criteria. However, EUNIS-ESy can identify more forest types (46), partly due to the use of herb-layer species in addition to trees in the definitions of individual habitats.

Other currently available pan-European expert systems were designed to identify phytosociological alliances or associations within a specific vegetation type, e.g., floodplain forests and alder carrs (Douda et al., 2016), beech forests (Willner et al., 2017), fens (Peterka et al., 2017), coastal dune grasslands (Marcenò et al., 2018), Mediterranean Lygeum spartum grasslands (Marcenò et al., 2019) and marshes (Landucci et al., 2020). Other expert systems have a more restricted geographic scope (see an overview with source codes at http://www.sci.muni.cz/botany/juice/?idm=25). These expert systems provided useful resources, and some species groups and decision rules proposed in some of them were used, with modifications, in EUNIS-ESy.

Some of the mentioned expert systems are designed to be applied only to the vegetation plots belonging to the broad habitat/vegetation type for which the expert system was developed. The plots not belonging to this scope have to be removed before classification; otherwise, they might be erroneously assigned to some of the types defined in the expert system. In contrast, EUNIS-ESy includes, in addition to the formal definitions of the coastal, wetland, grassland, shrubland, forest and man-made habitats, also preliminary definitions of all the other European vegetated habitat types. As a result, it can be applied to any vegetation plot from Europe.

Some expert systems for vegetation classification were also developed outside Europe. The expert system for national forest vegetation classification of Taiwan (Li et al., 2013), provided in a code executable in the R program, used a similar approach as the European expert systems, following the principles outlined by Bruelheide (1997, 2000) and Kočí et al. (2003). A different approach, based on supervised or semi-supervised fuzzy classification performed using the noise clustering algorithm, was applied for matching new plots to the units of existing national vegetation classification in New Zealand (Wiser and De Cáceres, 2013; Wiser et al., 2016).

4.3 Remarks on the practical application of EUNIS-ESy

4.4 Characteristic species combinations and distribution maps of habitats

Characteristic species combinations provided in this study are based on a statistical analysis of a large database of European vegetation plots, following the approach originally proposed by Chytrý and Tichý (2003) for an analysis of the Czech National Phytosociological Database and subsequently used for analyses of vegetation-plot databases in other countries (Jarolímek and Šibík, 2008; Kącki et al., 2013). The formal division of the characteristic species combination into diagnostic (concentrated in the habitat), constant (frequent, but not necessarily concentrated) and dominant (often attaining a high cover) provides a comprehensive characterization of each habitat through its plant species.

Species lists for European vegetation/habitat types are also provided in the electronic factsheets of the European Red List of Habitats (Janssen et al., 2016; https://forum.eionet.europa.eu/european-red-list-habitats/library/terrestrial-habitats) and the Supplementary material of EuroVegChecklist (Mucina et al., 2016, their Appendix S6; see also https://www.synbiosys.alterra.nl/evc/). However, both of these compilations are based on data from various sources and concepts developed independently by various experts, which introduces some inconsistencies. Moreover, these lists do not distinguish between diagnostic, constant and dominant species. In contrast, our lists are consistent across habitats, clearly discriminate the three categories of species included in characteristic species combination, and provide a numerical ranking of importance of each species within each category. Therefore, they can be used for various analyses and practical applications as the so far most reliable source of information on species composition of different European habitats. However, it is important to note that although we used a geographically stratified resampling of the data set, these species lists are, to some extent, biased due to considerable differences in vegetation plot density among European regions. In particular, species from northern and eastern Europe can be underrepresented in these lists. Moreover, information on bryophyte and lichen species is affected by the lack of their recording in many vegetation plots. Nevertheless, once the data sets from undersampled regions and with a more consistent recording of non-vascular plants become available, an extended data set can be classified by the current expert system and the species lists of characteristic species combinations can be updated.

The maps of habitat distribution based on the available European vegetation plots appear realistic for the habitats restricted to western, central and southern Europe, but have many gaps for the habitats occurring in or extending to the north and east. Distribution ranges of habitats can be successfully modelled if the original records of habitat occurrence cover a large part of the real range (Jiménez-Alfaro et al., 2018). However, our modelling exercise (Schaminée et al., 2014) yielded unstable and often incorrect results, especially in extrapolations to data-poor areas in eastern Europe. Therefore we refrained from complementing the maps with models here. Further work should be focused on identifying the areas with the most important data gaps for individual habitats and collecting data from such areas. The number of plots classified to each habitat reported in Table 1 partly reflects the occurrence frequency of each habitat in Europe, but it also reflects the research intensity. Special attention should be paid to the habitats that are so far represented by very few plots.