TITLE
Automated image processing for quantitative characterization of grassland vegetation structure: microhabitat selection in threatened meadow and steppe vipers
AUTHORS
Edvárd Mizsei1,2,3*, Mátyás Budai4, Gergő Rák4, Barnabás Bancsik5, Dávid Radovics1,3, Márton Szabolcs1, Attila Móré2,3, Csaba Vadász2, György Dudás6, Szabolcs Lengyel1
AFFILIATIONS
1Conservation Ecology Research Group, Department of Tisza Research, Danube Research Institute, Centre for Ecological Research, Hungary
2Kiskunság National Park Directorate, Hungary
3Department of Ecology, University of Debrecen, Hungary
4Department of Systematic Zoology an Ecology, Eötvös Loránd University, Budapest, Hungary
5Department of Ecology, University of Veterinary Medicine, Budapest, Hungary
6Bükk National Park Directorate, Eger, Hungary
*Corresponding author; e-mail: edvardmizsei@gmail.com
ABSTRACT
  1. Understanding animals’ selection of microhabitats is important in both ecology and biodiversity conservation. However, there is no generally accepted methodology for the characterisation of microhabitats, especially for vegetation structure.
  2. Here we present a method that objectively characterises vegetation structure by using automated processing of images taken of the vegetation against a whiteboard under standardised conditions. We developed an R script for automatic calculation of four vegetation structure variables derived from raster data stored in the images: leaf area (LA), height of closed vegetation (HCV), maximum height of vegetation (MHC), and foliage height diversity (FHD).
  3. We demonstrate the applicability of this method by testing the influence of vegetation structure on the occurrence of three viperid snakes in three grassland ecosystems: Vipera graeca in mountain meadows in Albania, V. renardi in loess steppes in Ukraine andV. ursinii in sand grasslands in Hungary.
  4. We found that the variables followed normal distribution and there was minimal correlation between those. Generalized linear mixed models revealed that snake occurrence was positively related to HCV inV. graeca , to LA in V. renardi and to LA and MHC inV. ursinii , and negatively to FHD in V. renardi , and to HCV in V. ursinii .
  5. Our results demonstrate that biologically meaningful vegetation structure variables can be derived from automated image processing. Our method minimises the risk of subjectivity in measuring vegetation structure, allows upscaling if neighbouring pixels are combined, and is suitable for comparison of or extrapolation across different grasslands, vegetation types or ecosystems.
KEYWORDS
biodiversity monitoring, ecological complexity, habitat diversity, habitat selection, reptile, Viperidae, visual obstruction reading
INTRODUCTION
Predicting the occurrence or abundance of animals hiding in the vegetation has been one of the earliest challenges for mankind and remains so for many ecologists. Understanding how animals choose microhabitats is a central aim in ecology and is fundamental for evidence-based conservation (Johnson et al., 2014). Habitat selection is a key evolutionary strategy because it both depends on and is influenced by resource availability and interactions with conspecifics and other species. Thus it has inevitable influence on individual fitness, and accordingly the evolution of life-history traits is associated with habitat properties (Morris, 2003). Habitat selection can thus be interpreted to reflect an adaptive strategy on evolutionary time scale optimalization (MacArthur et al., 1962; Pianka, 1973), although intraspecific competition and population density may also influence the choice of the individuals (Fretwell & Lucas, 1970; Lawlor & Smith, 1976), and these costs of a particular choice are rarely considered (Rosenzweig, 1981).
The characterisation of microhabitats, however, has proven to be difficult and there is no generally accepted methodology applicable across ecosystems, habitat types and animal groups (Stein et al., 2014). More complex habitats, i.e., those characterized by higher microhabitat diversity, are supposed to sustain a higher number of ecological niches and species occupying them compared to habitats with decreased structural diversity (MacArthur & MacArthur, 1961; Loke et al., 2015). Habitat or microhabitat diversity is often divided into two components: compositional diversity arises from the identity of different elements, whereas structural diversity arises from the two- or three-dimensional physical arrangement of the elements (Tews et al., 2004; Lengyel et al., 2016). Both aspects can be further subdivided into abiotic components (e.g., composition: soil types, hydrology; structure: elevation, topography) and biotic components (e.g., composition: plant species identity; structure: vegetation complexity). The quantification of abiotic elements and biotic compositional elements is usually straightforward via objective measurements (e.g., for soil types, hydrology: qualitative list of soil types, maps, measurement of groundwater table; for elevation, topography: GPS readings, landform diversity; plant species/association identity: list of species or plant associations). In contrast, a plethora of context-dependent methods have been used to measure vegetation structure (Mushinsky & McCoy, 2016). Several terms have been used for vegetation structure, such as structural complexity/diversity, canopy/foliage height/diversity, vegetation complexity/heterogeneity, architectural complexity (Tews et al., 2004). In studies of animal habitat selection, vegetation structure is often quantified by estimates of phytomass or by cover estimates. Additional methods include quantifying the presence or cover of structures formed by plants, e.g. tussocks, shrubs, dead phytomass such as leaves, height of shoots, leaf area, cover at various heights (Benkobi et al., 2000; Vermeire & Gillen, 2001; Pringle et al., 2003; Garden et al., 2007; Faria & Silva, 2010; Stumpel & van der Werf, 2012; Mizsei et al., 2020a). Many of these methods depend on subjective eyeball estimates made in the field confounded by observer bias and measurement error (Milber et al., 2008; Bergstedt et al., 2009), e.g. on plant cover, or return one value, e.g. vegetation height or phytomass, which, at most, is a proxy for the 3-D physical arrangement or distribution of vegetation elements. All these drawbacks prevent generalisations of animal-vegetation structure relationships across habitats, ecosystems and spatial scales. There is thus a clear need for objective methods that provide balanced measurements on multiple variables including both the horizontal and vertical distribution of vegetation elements and the one-value characteristics that succinctly summarise important aspects of vegetation structure.
Reptiles are among the most threatened vertebrates and decline globally due to habitat loss and degradation, introduced invasive species, environmental pollution, diseases, unsustainable use of natural/seminatural habitats and climate change (Gibbons et al., 2000). By now, one out of five reptile species has become threatened by global extinction, and local extinctions are becoming common. To cope with this, conservation priorities should be determined and actions should be implemented to reduce this rate (Böhm et al., 2013). In Europe, the meadow and steppe vipers (Vipera ursinii complex) are among the most threatened reptiles. Lowland populations of this complex (V. renardi , V. u. rakosiensis , V. u. moldavica ) lost almost all their habitats due to transformation of grasslands to croplands, and populations in Austria, Bulgaria and Moldova have gone completely extinct (Krecsák et al., 2003; Tupikov & Zinenko, 2015; Mizsei et al., 2018a). Alpine populations are threatened by overgrazing and climate change (Mizsei et al., 2020b). Although habitat restoration is increasingly used in reptile conservation (e.g. Péchy et al., 2015; Triska et al., 2016; Michael et al., 2018), little is known on the efficiency of these actions due to lack of knowledge on vegetation characteristics preferred by target reptiles or due to lack of proper monitoring (Block et al., 2001; Ruiz-Jaen & Aide, 2005; Jellinek et al., 2014; Mizsei et al., 2020b).
Here we present an approach to explore animal-vegetation relationships by objectively characterising vegetation structure by using photography-based standardized field data recording followed by computer-based automated quantification of particular structural attributes of the vegetation. We demonstrate the applicability of this approach in a case study using data collected in grassland habitats of three populations of the threatened V. ursinii complex of meadow vipers. While the applied field photography, based on images taken from vegetation against a whiteboard under standard conditions, has been used before, the algorithm-based quantification of vegetation structure, to our knowledge, is novel in the literature. We show that the variables derived from this methodology explain a significant part of the variation in snake occurrence and that the approach can thus be an important part of the repertoire of methods to characterise vegetation structure in studies of animal habitat selection.
MATERIAL AND METHODS