GRIMP: A machine‐learning method for improving groups of discriminating species in expert systems for vegetation classification

Abstract

Aims Expert systems are increasingly popular tools for supervised classification of large datasets of vegetation‐plot records, but their classification accuracy depends on the selection of proper species and species groups that can effectively discriminate vegetation types. Here, we present a new semi‐automatic machine‐learning method called GRIMP (GRoup IMProvement) to optimize groups of species used for discriminating among vegetation types in expert systems. We test its performance using a large set of vegetation‐plot records. Methods We defined discriminating species groups as the groups that are unique to each vegetation type and provide optimal discrimination of this type against other types. The group of discriminating species of each vegetation type considerably overlaps with the group of diagnostic species of this type, but these two groups are not identical because not all diagnostic species have sufficient discriminating power. We developed the GRIMP iterative algorithm, which optimizes the groups of discriminating species to provide the most accurate vegetation classification, using a training set of a priori classified plot records. We tested this method by comparing classification accuracy before and after the GRIMP optimization of species groups using vegetation‐plot records from the Czech Republic a priori classified to 39 phytosociological classes, and three initial sets of candidate discriminating species from different sources. Results The GRIMP algorithm improved the classification accuracy at the class level from 65% correctly classified plots in the test dataset before group optimization to 88% thereafter. The other plots were misclassified or unclassified, but misclassifications were reduced by adding further expert‐based criteria considering dominant growth forms. Conclusions GRIMP‐optimized groups of discriminating species are very useful for semi‐automatic construction of expert systems for vegetation classification. Such expert systems can be developed from an a priori unsupervised or expert‐based classification of at least some vegetation plots.

Classification of the European marsh vegetation (Phragmito‐Magnocaricetea) to the association level

“…Therefore we accepted only species dependent on these factors for distinguishing different associations dominated by the same species. We used the frequency of sociological species groups to assign the associations to the alliances. Then we made formal definitions of the high‐rank syntaxa a posteriori using a different approach than that used for defining associations (GRIMP method; Tichý et al, ). We used this procedure after several pilot attempts with other procedures including the creation of formulas for high‐rank syntaxa in the same way as we did for the associations.…”

“…A plot is assigned to the class Phragmito‐Magnocaricetea if the total cover of the ‘ Phragmito‐Magnocaricetea species’ is higher than the total cover of the ‘discriminating species group’ of any other class, or if the total cover of the ‘ Phragmito‐Magnocaricetea species’ is >25%. At the same time the total cover of ‘mire bryophytes’ or ‘shrubs and trees’ is not higher than 25% and the total cover of ‘other herbaceous plants and dwarf shrubs’ is not higher than 50%. Definition of orders — The ‘discriminating species groups of orders’ included in the class Phragmito‐Magnocaricetea were created using the semi‐automatic procedure called GRIMP (Tichý et al, ), which optimizes groups of species used for discriminating vegetation types in expert systems. A preliminary list of potentially discriminating species for each order was created by including all the species for which the association between the group of plots representing the order and the species occurrence across the total data set was significant at p < 0.001 (Fisher's exact test).…”

“…• Definition of orders -The 'discriminating species groups of orders' included in the class Phragmito-Magnocaricetea were created using the semi-automatic procedure called GRIMP (Tichý et al, 2019), which optimizes groups of species used for discriminating vegetation types in expert systems. A preliminary list of potentially discriminating species for each order was created by including all the species for which the association between the group of plots representing the order and the species occurrence across the total data set was significant at p < 0.001 (Fisher's exact test).…”

Oak-hornbeam forests of central Europe

“…Lists of the most frequent species in each group were compiled. Then we applied the GRIMP algorithm (Tichý et al 2019) to select among them the species that best discriminated each group from all the others, and such species were then used in the expert system. Afterwards, we applied this expert system to all the deciduous forest relevés from the above-mentioned sources.…”

Probabilistic key for identifying vegetation types in the field: A new method and Android application

“…The key question for the users of vegetation classification systems is how to identify vegetation types, either in the field or in databases of vegetation plots. Several methods have been developed for the identification of vegetation types in databases, some of them based on similarity in species composition between a single plot record and a set of plots previously classified to the types (Gégout & Coudun, 2012;Hill, 1989;Tichý, 2005;van Tongeren, Gremmen, & Hennekens, 2008), others based on expert systems comprising formal definitions of types (Bruelheide, 1997;Kočí, Chytrý, & Tichý, 2003;Landucci, Tichý, Šumberová, & Chytrý, 2015;Tichý, Chytrý, & Landucci, 2019). However, there are hardly any tools that would enable quick identification of vegetation types directly in the field.…”