{"id":120501,"date":"2025-10-14T02:15:13","date_gmt":"2025-10-14T02:15:13","guid":{"rendered":"https:\/\/www.europesays.com\/ie\/120501\/"},"modified":"2025-10-14T02:15:13","modified_gmt":"2025-10-14T02:15:13","slug":"learning-the-syntax-of-plant-assemblages","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/ie\/120501\/","title":{"rendered":"Learning the syntax of plant assemblages"},"content":{"rendered":"

A visualization of the methodology used in this paper is shown in Fig. 1<\/a>, a more complete overview is provided in appendix 26 in the Supplementary Information<\/a> and a detailed description of each step is shown in Supplementary Figs. 9<\/a>\u201311<\/a>. An explanation of all acronyms and terms can be found in Supplementary Texts 30<\/a> and 31<\/a>.<\/p>\n

Leveraging vegetation plots<\/p>\n

The data used for training the Pl@ntBERT model were extracted from EVA37<\/a>. EVA is a database of vegetation plots\u2014that is, records of plant taxon co-occurrence that have been collected by vegetation scientists at particular sites and times. The EVA data were extracted on 22 May 2023. They contained all georeferenced plots from Europe and adjacent areas (that is, 1,731,055 vegetation plots and 36,670,535 observations from 34,643 different taxa).<\/p>\n

These vegetation plots were first split into two sets, depending on the presence or absence of a habitat type label:<\/p>\n

    \n
  1. \n (1)<\/p>\n

    A dataset containing unlabelled data\u2014that is, vegetation plots with a missing indication of EUNIS habitat type. This dataset (henceforth \u2018fill-mask dataset\u2019) containing 572,231 vegetation plots could be used only for training the masked language model.<\/p>\n<\/li>\n

  2. \n (2)<\/p>\n

    A dataset containing labelled data\u2014that is, vegetation plots with an indication of EUNIS habitat type. This dataset (henceforth \u2018text classification dataset\u2019) containing 850,933 vegetation plots could be used for training both the masked language model and the text classification model.<\/p>\n<\/li>\n<\/ol>\n

    To ensure a clean dataset representing vegetation patterns well, some additional pre-processing steps were conducted. We removed the few species with a given cover percentage of 0, assuming these were errors or scientists reporting absent species (which resulted in 31,813,043 observations remaining). We merged duplicated species in the same vegetation plots (that is, species that appeared twice or more in one vegetation plot because they were in different layers) and summed their percentage covers (which resulted in 31,036,661 observations remaining). The taxon names were then standardized using the API of the Global Biodiversity Information Facility (GBIF). It relies on the GBIF Backbone Taxonomy as its nomenclatural source for species taxon names and integrates and harmonizes taxonomic data from multiple authoritative sources (for example, Catalogue of Life, International Plant Names Index and World Flora Online). As EVA is an aggregator of national and regional vegetation-plot databases, this step ensured that the same species collected in two very distant areas still shared the same name67<\/a>. If no direct match was found for the species name (for example, the GBIF Backbone Taxonomy was not able to provide a scientific name for the EVA species Carex cuprina), then it was dropped. As we focused on the species taxonomic rank, taxa identified only to the genus level were dropped, and taxa identified at the subspecies level were lumped together at the species level (for example, Hedera was dropped but both Hedera helix subsp. helix and Hedera helix subsp. poetarum were merged into Hedera helix). This resulted in 29,859,407 observations remaining. We removed hybrid species and very rare species (that is, species that appeared less than ten times in the whole dataset), which resulted in 29,836,079 observations remaining. Vegetation plots that lost more than 25% of their taxa or their most abundant taxon after the species name matching were removed from the dataset, to ensure that the remaining plots still provided reliable representations of vegetation patterns (which resulted in the final number of 29,149,022 observations remaining). Finally, vegetation plots belonging to very rare habitat types (that is, habitat types that appeared less than ten times in the whole dataset) were considered unlabelled data and added to the fill-mask dataset.<\/p>\n

    Fig. 5: Overview of the framework.<\/b>\"figure<\/a><\/p>\n

    The sequence of tasks performed during each of the five main stages (installation check, dataset curation, masking training, classification training and outcome prediction).<\/p>\n

    The set of labelled vegetation plots was then strategically split. To avoid overfitting, ideally part of the available labelled data must be held out as a test set. However, the quantity of available full lists of plant species with estimates of cover abundance of each species and habitat type assignment is not very high (that is, less than 1,000,000 vegetation plots for all of Europe, a relatively low number compared with the vast amount of biodiversity data available). Partitioning the available data into a training set and a test set would reduce the number of training samples to a level too low for effective model training. It is therefore possible to use k-fold cross-validation to split the dataset into k subsets instead. Then, for each of the splits, the model can be trained using k\u2009\u2212\u20091 of the subsets for training and the latter one for validation. However, cross-validation scores for the classification of vegetation plots can be biased if the data are randomly split, because they are commonly spatially autocorrelated (spatially closer data points have similar values). One strategy to reduce the bias is splitting data along spatial blocks68<\/a>. This procedure avoids fitting structural patterns and allows the separation of near-duplicates. Such vegetation plots differ from each other in a very small portion of species (for example, if they are close in space, two vegetation plots may exhibit identical plant composition but feature species with slightly contrasting abundances). The dataset was thus first split into spatial blocks of 6\u2009arcmin (0.1\u00b0 on the World Geodetic System 1984 spheroid). The blocks were then split into folds. Since the geographic distribution of vegetation plots across Europe is unequal, each block can have a different number of data points. The folds were thus balanced to have approximately equal numbers of plots instead of assigning the same number of blocks to each fold (which could have led to folds with very different numbers of data points). This process was facilitated by the use of the research software Verde.<\/p>\n

    With over 1,400,000 vegetation plots, 29,000,000 observations and 14,000 species, the dataset used in this paper is one of the most extensive datasets of vegetation plots ever analysed69<\/a>. The entire description of the dataset can be found in Supplementary Table 2<\/a>, and a visualization of the data can be found in appendix 32 in the Supplementary Information<\/a>. An overview of the long-tail distribution of species (that is, there is a strong class imbalance, meaning that a few species are present in many of the vegetation plots) can be found in Supplementary Fig. 14<\/a>, and more taxonomic information on the species (for example, class, order and family), mostly vascular plants with some bryophytes and lichens, can be found in appendix 16 in the Supplementary Information<\/a>.<\/p>\n

    The EUNIS habitat types18<\/a> are referred by their codes instead of their names, as they better reflect the classification hierarchy. The coding system is structured so that each broad habitat group is represented by one letter (except the broad habitat group littoral biogenic habitats, which is designated by the code MA2). A new alphanumeric character is then added for each subsequent level. For instance, the habitat type Mediterranean, Macaronesian and Black Sea shifting coastal dune is identified by the code N14, indicating its belonging to the habitat group N1 (that is, coastal dunes and sandy shores), and more generally to the broad habitat group N (that is, coastal habitats). The entire list of the 227 habitat types used in this work can be found in appendix 24 in the Supplementary Information<\/a>, but to exemplify the habitat types included, we list the eight broad habitat groups used in this paper below:<\/p>\n