We built indicators of grassland type suitability as habitat for butterflies, bumblebees, and grasshoppers using information from the scientific literature and expert interviews. Grassland management is known to have a strong impact on insect abundance and species richness; therefore, management had an important weight in our indicators. Management was defined according to the defoliation regime (cutting, grazing, or mixed) and the intensity of defoliation, soil fertility, and timing of the defoliation regime. Soil fertility was estimated from fertilizer application, and we used a 50 kg.ha⁻¹.year⁻¹ threshold for mineral and organic fertilization to differentiate between low and high fertility. We proposed 16 management types (Table 1) that can be easily identified by researchers or stakeholders in most situations. The effect of each management type on butterflies, bumblebees and grasshoppers was evaluated through their effect on sward surface height and sward height variability.

Sward surface height (SSH) is indeed crucial for butterflies and grasshoppers, as biomass accumulation provides them with more food resources, as well as suitable shelters and favorable microclimates. Butterfly abundance and species richness have been positively correlated with vegetation height (Kruess and Tscharntke, 2002; Pöyry et al., 2006; Sjödin et al., 2008; Scohier et al., 2013; Larkin and Stanley, 2021). In addition, sward height variability is likely to benefit butterfly diversity because (1) butterfly assemblages vary according to vegetation height (Pöyry et al., 2006; Sjödin et al., 2008; Jerrentrup et al., 2014; Van Halder et al., 2017), and (2) some species require short and tall vegetation at different stages of their developmental cycle (Dennis, 2004; WallisDeVries, 2004). For grasshoppers, Marini et al. (2009) concluded that sward structure was more important than food quality in managed grasslands. Grasshopper species richness has been positively correlated with sward height variability (Kruess and Tscharntke, 2002; Fumy et al., 2020), and there are some well-documented surveys of grasshopper species relying on a mosaic of short and tall patches at different stages of their developmental cycle (Cherril and Brown, 1992; Willott, 1997). Grasshopper abundance is also usually higher in tall and extensively managed swards (WallisDeVries et al., 2007; Marini et al., 2008, 2009; Dumont et al., 2009; Kati et al., 2012; Fleurance et al., 2016).

Bumblebees were assumed to respond differently to SSH since many bumblebee species have a habitat preference for open grasslands (Williams and Osborne, 2009). However, contrasting sward height preferences have been reported for different bumblebee species (Carvell, 2002; Sjödin, 2007), which suggests there is a benefit of increasing sward height variability on bumblebee species richness (Scohier et al., 2013; Ravetto Enri et al., 2017). SSH is also a driver of bumblebee abundance since keeping grass short severely affects the survival of surface nests. Members of Bombus pascorum that nest among grass more than other species preferentially used tall grasses rather than patches where grass was frequently defoliated (Fussell and Corbet, 1992). The litter must also not be too thick so that bumblebees can dig their nests (Carvell, 2002).

Beyond the effects of sward structure on insect communities, sward botanical composition is crucial for flower-visiting insects (Berg et al., 2019). Butterfly species richness has been shown to be positively correlated with plant species richness (Öckinger et al., 2006; Skórka et al., 2007; Marini et al., 2009). Grasslands with higher plant species richness indeed provide host plants for more butterfly species and offer a nectar source that is better distributed in time compared to species-poor grasslands. This is assumed to benefit a significant part of butterfly assemblages (Jeanneret et al., 2003a,b; Van Halder et al., 2017). Dry and nutrient-poor grasslands belong to the most species-rich habitats for butterflies and especially for endangered ones in Europe (Van Swaay et al., 2006) suggesting that site temperature and nutrient status influence butterfly species richness. However, some of these relationships differ across European biogeographical areas and were therefore not considered. One example is temperature, which is usually positively correlated to butterfly populations (e.g., in Britain; Turner et al., 1987), while butterfly species richness was negatively correlated with temperature in Catalonia (Stefanescu et al., 2004). Butterfly abundance has been consistently positively related to the abundance of plants providing nectar across Europe (Bergman et al., 2008; Farruggia et al., 2012; Scohier et al., 2013; Jerrentrup et al., 2014). Alison et al. (2021) recently confirmed this positive relationship, while pointing out the saturating benefit of flower cover on butterfly abundance. Habitat characteristics influencing the abundance, foraging activity and species richness of bumblebees also include the diversity and abundance of flowering plant species (Carvell, 2002; Walcher et al., 2019; Larkin and Stanley, 2021; Phelan et al., 2021). Legumes appear to be the major pollen source for most bumblebee species (Goulson et al., 2005; Williams and Osborne, 2009; Scohier et al., 2013; Marja et al., 2021); thus, legumes strongly affect bumblebee abundance. Other entomophilous plant species such as Centaurea and Cirsium spp. were also reported as a key pollen source for bumblebees (Scohier et al., 2013; Marja et al., 2021; Phelan et al., 2021).

Finally, Walcher et al. (2019) found a significant positive correlation between Orthoptera and vascular plant species richness in semi-natural grasslands of the Austrian Alps. Additionally, Kati et al. (2012) concluded that low disturbance intensity, as well as plant species richness, predict grasshopper species richness in humid Mediterranean grasslands. They even suggested the use of grasshoppers as surrogates for vascular plants and vice-versa, given their congruent species richness patterns. Tall and dense swards may lead to lower temperatures in the egg environment, which can negatively affect thermophilous species and therefore Orthoptera abundance (Van Wingerden et al., 1991; Gardiner, 2006). The Ellenberg temperature index was thus also used as variable for evaluating effects on grasshopper species richness and abundance.

This literature review, together with expert interviews, was used for assessing the relative weight of each variable used in multicriteria decision trees as described in the following section.

Our composite indicators were based on decision trees created with fuzzy inference software called FisPro (Guillaume and Charnomordic, 2011). FisPro allows for the creation of a decision tree in which each node is a fuzzy inference system (Guillaume and Charnomordic, 2012). Fuzzy partitioning makes it possible to account for uncertainty in the choice of threshold values between the different alternatives (Figure 1). In fuzzy logic, there is a continuous evolution between truth and falsity, which makes a more precise assessment possible compared to the classical logic, where propositions such as “butterfly abundance is low” or “butterfly abundance is high” are either true or false according to a threshold value. Fuzzy logic defines membership functions in order to map continuous variables into classes such as “high” and “low.” Therefore, each abundance level has a degree of membership to both the “high” and “low” classes. The membership function shape can be parametrized based on existing knowledge, and a linear relationship is usually chosen when no information is available. This method softens the transition between classes compared to classical logic. The mapping between input and output variables is encoded through a set of decision rules, which can include rules such as “if input1 is low and input2 is low, then output is low” or “if input1 is high and input2 is low, then output is high.” Since input1 belongs to both the low and high classes, and has a given degree of membership for each, both decision rules are triggered and are weighted by the input1 and input2 membership degrees. As a result, the output includes a membership degree for each output classes. A final step, called defuzzification, transforms these membership degrees into a single numerical value (Suárez and Lutsko, 1999; see Figure 1).

Hierarchical decision trees are based on cause-effect relationships and, therefore, allow the implementation of several key components and hierarchical steps; each node corresponds to one of the key components, and each branch presents a relationship between two key components. The different nodes and hierarchical steps provide structure to implement the input, intermediate and output variables. The example reported in Table 2 details how bumblebee species richness was predicted. According to Section Sward Variables Affecting Insect Diversity, bumblebee species richness was assumed to rely equally on two criteria: species richness of entomophilous plants and variability in SSH. The result of the decision trees was a rough membership degree of the three classes, i.e., unfavorable, favorable, or highly favorable to bumblebee species richness. After the defuzzification process performed by FisPro (Figure 1), these discrete categories were transformed into a continuous value between 0 (null) and 10 (excellent habitat value for bumblebees), as shown in Figure 4.

Thus, the indicators are based on relatively simple decision trees that account for key variables that affect the abundance and species richness of each of the three insect taxa, as reviewed in Section Sward Variables Affecting Insect Diversity. The relative weight of each node was based on the literature review rather than on the correlative relationships between vegetation and insect assemblages that have been quantified in other articles (e.g., Jeanneret et al., 2003a; Bergman et al., 2008); such correlative relationships were considered too site-specific and can differ across sites (e.g., how temperature affects butterfly species richness: Turner et al., 1987; Stefanescu et al., 2004). The search for parsimony led us to suppress some variables from the decision trees using an iterative process; this was deemed acceptable as long as tree predictive power was not degraded. The calculation of indicators (Tables 2–5) requires three types of information: (1) information on grassland management (e.g., livestock unit per ha, number of cuts, general level of fertilization), (2) information from the floristic surveys (e.g., plant species richness, legume abundance) and (3) information calculated from the survey (e.g., abundance and species richness of entomophilous species, Ellenberg Temperature Index). Consulting the eFLORAsys database (http://www.eflorasys.univ-lorraine.fr) makes it possible to determine for each species, its Ellenberg temperature index, and whether or not it is entomophilous species. The Ellenberg index is calculated by taking the average of the values of the indices of the species present.

Parameter files and instructions (see README.txt file) for the effective computation of the indicators using FisPro are available online at https://doi.org/10.15454/A7KNLK (Dumont et al., 2022). This requires a complete list of plant species and an estimation of the weight contribution of each species at the plot level.

According to Bockstaller and Girardin (2003), there are three different ways to validate an indicator: the “design validation” is used to evaluate if an indicator is scientifically sound, the “output validation” is used to assess the soundness of the indicator outputs, and the “end-use validation” is used to ensure that the indicator can be used as a decision tool. Design validation requires the assessment of the scientific quality of indicator construction. In our case, the construction of multicriteria decision trees was based on the literature review and additional expert interviews. It required frequent feedbacks in order to identify variables that did not influence the decision-tree outputs after they were run using a set of experimental data. In this case, specific papers pointing out the effect of such variables were read again to re-evaluate any misleading interpretation and to reassess the variable weight in the decision tree.

Most of the validation process was based on the soundness of indicator output, or the so called “output validation.” The outputs from the decision trees were then compared to experimental data from the five following projects:

Discrepancies between datasets were solved by selecting surveys that used the same protocol; all adult insects were counted using “Pollard walks” along 50-m-long transects under good weather conditions (Pollard and Yates, 1993). The overall sampling effort varied across experimental sites; there were either two or three transects per paddock, and the number of measurement days between April and September was up to 10 days per year. Standardized abundances were obtained by selecting the subset of data within the 3 months that had the highest recorded abundance (usually June-August). Minimum of counts during this period was for grasshoppers in S3 that were only recorded on two occasions (Dumont et al., 2009), but fewer counts than for butterflies were judged adequate in view of the longer life span of grasshoppers (WallisDeVries et al., 2007). Individual counts were then averaged along each transect in each paddock. Thus, standardized abundance refers to the number of individuals encountered along a 50-m-long transect. Standardized species richness was obtained by calculating Chao estimates from raw observation data (this index extrapolates total species richness by estimating the number of unseen species) using the “vegan” package for R (Oksanen et al., 2013).

Output validation was based on a graphical test that checks whether, after defuzzification, decision trees predict a range of contrasting biodiversity outputs (on the X-axis) and whether these predicted data are consistent with the standardized observed data (on the Y-axis). The observed diversity can be lower than the predicted habitat suitability potential due to unexpected limiting factors (Cade and Noon, 2003), including unfavorable climatic conditions (Turner et al., 1987; Henry et al., 2015). The opposite, i.e., high observed data in poorly rated grasslands, should not be obtained. Although the data we used for output validation came from sound experiments, we did not expect a 1:1 linear relationship between the indicator outputs and observed data. Rather, Mitchell (1997) and Girardin et al. (1999) proposed a graphical test based on a “probability area.” Here, the probability area was defined as the triangle enclosed between the X-axis and a slope line that matched the minimum observed value to an indicator value of 0 and the maximum value to 10. One index point was added to this slope line as an envelope of acceptable prediction risk (see the white area in Figures 2, 4, 5). The probability area is assumed to include at least 95% of the points. By doing this, the probability of rejecting indicator validation (p-value) is estimated from the proportion of points outside the area (Bockstaller and Girardin, 2003).

The butterfly species richness indicator was then calculated using an independent dataset of 395 French grasslands; this dataset included a wide range of pedoclimatic conditions and an altitudinal gradient from Atlantic lowlands to upland pastures (see Plantureux and Thorion, 2005; Michaud et al., 2012 for plot location and observation method). This calculation was done to illustrate indicator sensitivity. In the independent dataset, plant species richness ranged from one to 69 species per plot, and grasslands were distributed across the 16 management types reported in Table 1. The differences in indicator values between management types were assessed with a Kruskal-Wallis test followed by a post-hoc test for pairwise comparison (Siegel and Castellan, 1988).

Based on the literature review and expert interviews, we propose the decision tree in Table 3 to predict the effects of management and botanical composition on butterfly abundance and species richness. Figure 2 compares the outputs of the butterfly decision tree (i.e., an indicator value ranging from 0 = unfavorable to 10 = highly favorable grassland for butterfly species richness and abundance) after defuzzification (on the X-axis) and the standardization of recorded data from the eight experimental datasets across Germany, France, Switzerland, and Italy (Y-axis). The outputs of the butterfly decision tree fit the observed data for species richness well, since data could be plotted within the triangle-shaped probability area (P < 0.05; Figure 2). The output validation was more difficult for butterfly abundance in spite of the tendency for data to be plotted within the triangle (P = 0.10). Most of the points outside the triangle came from the same survey in Germany.

The sensitivity test of the butterfly species richness indicator to grassland management using the independent dataset of 395 grasslands detected contrasted suitability of grasslands, with consistent and significant differences (χ²₁₅ = 121.16; P < 0.05) among the 16 management classes (Figure 3). In high-yield cut grasslands, the indicator predicts that butterflies would benefit from lowering the cutting frequency, but the median indicator values were not sensitive to the number of cuts in low-yield grasslands (i.e., only variability decreased as the number of cuts increased). Continuously grazed pastures with a lenient stocking rate appeared to be more suitable for butterflies than pastures grazed with a high stocking rate. Conversely, there was no significant effect of the rotation type on the indicator value, which was significantly lower than the effect found in continuously grazed plots with the lowest stocking rate. Additionally, having either one or two cuts before grazing did not affect the indicator value for butterfly species richness. A detrimental effect of management was, however, observed in grasslands that were cut before being rotationally grazed under short rotation, i.e., an intensive management that strongly limited sward heterogeneity.

We propose the decision tree from Table 4 to predict the effects of grassland management and botanical composition on bumblebee abundance and species richness. Figure 4 compares the decision tree outputs (indicator value ranging from 0 = unfavorable to 10 = highly favorable grassland for bumblebees) after defuzzification (on the X-axis) and standardization of the data recorded from the five experimental datasets across Wales, France and Switzerland (Y-axis). The outputs of the bumblebee decision tree fit the observed data of bumblebee abundance (P < 0.05) and species richness (P < 0.01), since data could be plotted within the triangle-shaped probability area (Figure 4).

We propose the decision tree from Table 5 to predict the effects of management and botanical composition on grasshopper abundance and species richness. Figure 5 compares the decision tree outputs (on the X-axis) and standardization of the data recorded from the four experimental datasets across Germany, France and Italy (Y-axis). The outputs of the grasshopper decision tree fit the observed data of grasshopper abundance (P < 0.05) and species richness (P = 0.05), since data could be plotted within the triangle-shaped probability area. All the points outside the triangle came from the same diverse grassland community on shallow soil from the Italian survey. The range of indicator values was also more narrow in the case of grasshopper abundance than in all other cases.