Translating a semi-supervised classification into a hierarchical expert system: Navarran grasslands demonstrate a way forward in classification approaches
The post provided by Idoia Biurrun & Jürgen Dengler
This post refers to the article Grasslands of Navarre (Spain), focusing on the Festuco–Brometea: classification, hierarchical expert system and characterisation by García-Mijangos et al. published in Vegetation Classification and Survey (https://doi.org/10.3897/VCS/2021/69614).
Classification of plant communities is important for organising our knowledge on the diversity of vegetation cover and its relationship to environmental gradients. It is crucial for biodiversity conservation, as it allows selecting the appropriate management measures depending on vegetation type. This is clearly applicable for grassland habitats, especially if we take into account that in broad areas these are secondary grasslands, which can be very diverse but depend on human intervention for their conservation. This is the case of the Spanish region of Navarre.
Located in the northern-central Iberian Peninsula, Navarre is a diverse region, both in terms of biogeography and environmental conditions, as strong geologic, topographic and climatic gradients can be found across the 10,391-km2 large territory. Natural vegetation is mostly forest, except for small areas in the alpine zone in the Pyrenees and in drylands of Ebro valley. Nevertheless, secondary grasslands such as meadows, dry grasslands and Mediterranean grasslands, occupy big extensions throughout the region.
In view of the strong environmental gradients present, high grassland diversity is expected in Navarra. One of the co-authors of this paper, Asun Berastegi, focused her PhD thesis on the survey and classification of Navarran grasslands. She performed a thorough sampling balanced across grassland types and subregions, with a narrow range of plot sizes, many of them with a fixed size of 3 m x 3 m. She distinguished 69 grassland associations or communities belonging to 11 phytosociological classes (Berastegi 2013), but also concluded that the delimitation of grasslands in the submediterranean areas still needed clarification and that the interpretation of the classes Ononido-Rosmarinetea and Festuco-Ononidetea was unclear.
Encouraged by this high grassland diversity and by the aforementioned delimitation issues, the Eurasian Dry Grassland Group (EDGG; https://edgg.org/) organized a Field Worskhop in Navarre in 2014 (Biurrun et al. 2014) with the aims to clarify grassland classification and to assess alpha diversity and fine-grain beta diversity of these grasslands by means of the standardized EDGG sampling, which includes bryophytes and lichens as well as in situ measurements of topographic and structural parameters and collection of soil samples for further analysis of soil parameters (Dengler et al. 2016). 35 biodiversity plots and 49 additional normal plots were surveyed during the Field Workshop.
In this study, we combined both datasets to classify the Navarran grasslands numerically and to translate the results into a syntaxonomic classification system. 839 plots with areas ranging from 5 to 25 m2 (mostly 9 m2) extracted from Berastegi (2013) were combined with 119 10-m2 plots sampled during the Field Workshop, to form a dataset of 958 plots, 119 of them also including bryophytes and lichens as well as soil, structural and topographic parameters.
Methodologically, we built on various contemporary approaches and combined them into a hierarchical workflow. At each hierarchical step (all grasslands to classes; class to orders; order to alliances; alliance to associations) we started with an unsupervised TWINSPAN classification whose clusters were sometimes slightly modified, e.g. merged when they contained very few relevés. Then we determined for each modified cluster diagnostic species based on phi-coefficients (Tichý & Chytrý 2006). These diagnostic species were used without modification to construct an electronic expert system in the Cocktail language of JUICE (Tichý 2002; Chytrý 2007; Chytrý et al. 2020). This electronic expert system was then applied to the relevés again, leading to moderate changes in the group assignments that sharpened the floristic division between the units. Finally, diagnostic species were determined again for the refined units, each of which was then submitted to the same procedure at the next hierarchical level.
As a result, we could present a hierarchical expert system that is able to classify close to 100% of all plots in a meaningful way. This contrasts with the “traditional” Cocktail approach that typically leaves 50–70% of all plots unclassified (Chytrý 2007). This much higher classification rate is important when it comes to the practical application of a classification system, e.g. in vegetation mapping. Two aspects contributed to the higher classification rate: (a) assignment of plots to the unit whose diagnostic species were prevailing (similar to Chytrý et al. 2020); (b) hierarchical top-down structure of the expert system. Another difference, and in our view an advantage, of our approach to create the expert system is that we first “let the data speak”, i.e. started with an unsupervised classification from which we derived diagnostic species, while in current literature expert systems are often artificially created based on expert knowledge who aim to mimic phytosociological concepts from the literature. Starting the classification with a large and balanced dataset enables finding overlooked, misplaced and dubious entities from traditional classifications instead of “forcing” the expert system to reproduce them. In the appendices of our paper, we provide the electronic expert system so that it can be run in JUICE to classify any future relevé into our proposed classification system. We used our expert system also to classify type relevés of relevant syntaxa, which enabled us to determine the correct name of each of our associations, i.e. the oldest validly described association name whose type relevé was assigned to this unit (Theurillat et al. 2021). Since the Cocktail language currently does not support the automated top-down classification in a hierarchical manner, for the time being one needs to run it in four subsequent steps, but we hope that in the future such a hierarchical approach can be directly reflected in the Cocktail syntax.
In this initial study, we applied our approach to delimit the classes within all grasslands and then picked the class Festuco-Brometea (basiphilous dry grasslands) to classify it down to association level. While our study very well supported the eight associations previously delimited by Berastegi (2013), it suggests some important modification of the syntaxonomic system at higher level. Navarran submediterranean grasslands previously included in the classes Festuco-Ononidetea and Ononido-Rosmarinetea are now included in the Festuco-Brometea, as two alliances forming a new xeric order (hitherto unnamed) and opposed to the two alliances of the meso-xeric order Brachypodietalia pinnati. On the other hand, subalpine grasslands previously included in the Festuco-Ononidetea were classified in the Elyno-Seslerietea. Thus, we could not define a separate class that could be interpreted as Festuco-Ononidetea in the Navarran context. While we could not use bryophytes and lichens to construct the expert system as they only had been determined in the subset of plots from the EDGG Field Workshop, we added them posthoc and determined diagnostic species among them, using only the plots within which they had been searched for. This resulted in numerous diagnostic cryptogams, particularly in the Mediterranean grasslands where they had been largely neglected in traditional phytosociology.
While this study contributed one important step in the classification of grasslands in Navarre and the Northern Iberian Peninsula, we refrained from formally describing the new syntaxonomic units at order and alliance level, as the area covered by our study was not big enough. Similarly, the decision whether the Festuco-Ononidetea should be kept as a separate class as in Mucina et al. (2016) or dissolved as our data suggest, needs further analyses covering broader geographic and ecological extents, including the most typical stands of this class. With this idea in mind, some of the co-authors of this paper have initiated a new project aiming to classify the basiphilous perennial dry grasslands and scrubs across all of Southern Europe.
Brief personal summaries:
Idoia Biurrun is an Associate Professor at the University of the Basque Country UPV/EHU in Spain. She organized the 7th Field Workshop in Navarre and is now deputy coordinator of these research expeditions in the Eurasian Dry Grassland Group and database manager of the GrassPlot database that makes all these data available. She is interested in the classification of plant communities and in many aspects of vegetation diversity and ecology.
Jürgen Dengler is a Professor of Vegetation Ecology at the Zurich University of Applied Science (ZHAW) in Wädenswil, Switzerland. He co-founded the Eurasian Dry Grassland Group (https://edgg.org/) and for a long time coordinated its Field Workshops. His major research interests are grassland ecology, grassland conservation, biodiversity patterns, macroecology, vegetation change, broad-scale vegetation classification and methods development in vegetation ecology.
