An objective- based prioritization approach to support trophic complexity through ecological restoration species mixes

1. Reassembling ecological communities and rebuilding habitats through active restoration treatments require curating the selection of plant species to use in seeding and planting mixes. Ideally, these mixes should be assembled based on attributes that support ecosystem function and services, promote plant and animal species interactions and ecological networks in restoration while balancing project constraints. Despite these critical considerations, it is common for species mixes to be selected opportunistically. Reframing the selection of seed mixes for restoration around ecological objectives is essential for success but accessible methods and tools are needed to support this effort. 2. We developed prioritizing plant species attributes to best support different objectives for ecosystem functions, services and trophic relationships such as pollination, seed dispersal and herbivory. We compared results to approaches where plant species are selected to represent plant taxonomic richness, dominant species and at random. We tested our framework in European alpine grasslands by identifying 176 plant species characteristic of the species pool, and identified 163 associated attributes affiliated to trophic relationships, ecosystem functions and services. In all cases, trophic can be captured based prioritization using the functional identity of plant species. (plant species lists) can be quantita-tively, in of costs, species or We confirm that a random draw of


| INTRODUC TI ON
The specific objectives of terrestrial ecological restoration will vary, but generally aim to return a habitat to a naturally functioning and stable state. Restoration is often operationalized through the planting or seeding of mixtures of plant species as an active treatment meant to re-establish plant communities in degraded sites, usually informed by the plant species composition at reference sites (Brudvig & Mabry, 2008;Zobel et al., 1998). This begins by defining a species pool for the ecosystem of interest (the regional species pool, Zobel et al., 1998), and by taking stock of what species can be sourced from the wild or from commercial native seed producers (the restoration species pool; Ladouceur et al., 2018;Zobel et al., 1998).
However, seed and planting mixes used for restoration are often a low-diversity subset of the relevant species pool and are composed opportunistically (Barr et al., 2016). Species selection must be balanced within project constraints (e.g. budgets, labour, time) and within other project targets (e.g. increase plant cover, prevent erosion). These species mixes have a major impact on restoration success and have implications for the multi-taxa functionality of the restored ecosystems (Guiden et al., 2021). How species mixes can be optimized to maximize restoration goals efficiently within project constraints remains an open question and an urgent task for implementing the United Nations Decade on Ecosystem Restoration.
It is widely recognized that rebuilding habitats requires the consideration of ecosystem services and functions, fauna and plantanimal relationships (Kollmann et al., 2016;McAlpine et al., 2016).
Integrating these relationships in ecological restoration is a complex task that remains largely unaddressed despite increased calls for consideration (Cross et al., 2020;Dixon, 2009;Lindell, 2008;Majer, 1989Majer, , 2009Menz et al., 2011). Plant functional traits can help identify ecosystem services or functions facilitated by plant species, and optimizing functional diversity or particular trait convergence in restoration species selections has been shown to lead to favourable outcomes (Brudvig & Mabry, 2008;Laughlin et al., 2018;Wang et al., 2020).
Fauna also contribute crucial ecosystem functions to plants, such as seed dispersal (regeneration), pollination (seed production), herbivory (reduction of competitive dynamics) and patchy nutrient return (Olff & Ritchie, 1998). Optimizing plant species mixes to facilitate multiple ecosystem services and functions, including those facilitated through trophic relationships, could thus enhance restoration success. This has been demonstrated in the establishment of fruit-bearing trees to facilitate dispersal from other diverse patches by frugivores in tropical rainforests (Heelemann et al., 2012;Lamb et al., 1997), but relationships like these are important in other habitat types as well, and remain underexplored in restoration ecology.
However, the restoration of ecosystem services, functions and animal communities simultaneously is challenging due to complex processes, life cycles and dependence on plants as well as other trophic levels (Chan et al., 2006;Guiden et al., 2021). Plant-animal interaction networks, both mutualistic (pollination and frugivory) and antagonistic (herbivory), are highly non-random Lewinsohn et al., 2006;Rezende et al., 2007), and a disruption in these interactions can lead to trophic cascades across and within systems (Knight et al., 2005;Valiente-Banuet et al., 2015). Plant-animal interaction networks are often nested, that is, some species have many interactions in their networks, and many species have few (Bascompte et al., 2003). When considering balancing project constraints and restoration targets in a relatively low-diversity species mix for restoration, it is unlikely that a random draw from the regional plant species pool will provide resources to optimize trophic networks and other ecosystems services and functions. Systematic decision-making can quantitatively support complex multivariate decision-making problems such as this (Chan et al., 2006;Hill et al., 2014;M'Gonigle et al., 2016), with extremely flexible and diverse potential applications.
Here, we present a proof-of-concept for the optimization of active restoration species mixtures (for seeding or planting treatments) for supporting different objectives. We used species-rich European plant species from the regional plant species pool cannot be assumed to support other trophic groups and ecosystem functions and services.

Synthesis and Applications.
Our framework is presented as a proof-of-concept to help restoration practitioners better apply quantitative decision support to plant species selection to efficiently meet ecological restoration outcomes. Our approach may be tailored to any restoration initiative, habitat or restoration targets where seeding or planting mixes will be applied in active treatments. As global priority and resources are increasingly placed into restoration, this approach could be advanced to help make efficient decisions for many stages of the restoration process.

K E Y W O R D S
decision support, functional traits, optimization, plant traits, seed mixes, species selection, trophic networks subalpine and alpine calcareous grasslands as a case study. These habitats are sensitive to disturbance, and impacted by ski resorts and other tourism activities, making them a target system for ecological restoration across European Natura 2000 sites (Garcia-Gonzalez, 2008). We identified 176 plant species that frequently occur in the target ecosystem on a biogeographical scale as the potential regional and restoration species pool of interest (Ladouceur et al., 2018;Zobel et al., 1998). We used trait databases and literature to compile traits related to regeneration and relationships between the 176 plant species in our species pool and the insects, birds and mammals that are typical of these habitats and depend on particular plant species for various life stages. Hereafter, we refer to the traits and aspects of plant species that represent these relationships and characteristics of interest, as plant attributes.
Our primary aim is to develop and evaluate a quantitative decision-making framework to assist in species selection for seeding and planting mixes for restoration projects. To do so, we designed five objectives for prioritizing plant attributes that support ecosystem functions, services and trophic dependencies. We optimized for these objectives by finding the smallest number of plant species needed to deliver all of the attribute targets set within each objective (Possingham et al., 2000). We then developed four plausible species selections to compare with prioritized selections including a focus on dominant species, random draws from the plant species pool to represent different taxonomic resolutions, and completely random draws from the plant species pool. We used our study system to investigate whether optimized species pools deliver objectives for ecosystem functions or services more efficiently than selecting dominant plant species, for taxonomic richness, or randomly.

| MATERIAL S AND ME THODS
We designed and tested an optimization approach to prioritize species mixes for planting or seeding in restoration projects based on ecological objectives. Below, we describe how we (a) selected plant species that represent a defined regional species pool; (b) identified plant attributes that contribute to different restoration outcomes; (c) constructed objectives and optimized attributes and (d) evaluated across optimized plant species mixes based on objectives.

| Species selection
We compiled a list of the most frequent native species occurring in alpine calcareous grassland habitat types on a continental scale, using a synthesis of >1 million field surveys , reporting species frequencies in the habitat types of the European habitat classification system (EUNIS, https://www.eea.europa.eu/ data-and-maps/data/eunis -habit at-class ifica tion), directly assigned to habitat types of conservation concern (see Table S1). We identified native plant species that occur above a particular frequency (>5% of total occurrences) in calcareous alpine grassland habitat types on a European-wide scale. Expert opinion suggests that species below this frequency were found to be more typical of other habitat types. This resulted in a list of 176 native plant species that occur frequently in the calcareous alpine grasslands of continental Europe. We considered this to be the species pool of this habitat and we assumed all species can co-occur or can coexist. Furthermore, we consider all plant species in the pool as equal candidates for inclusion in seed mixes to meet prioritization objective targets.

| Attribute selection
For the 176 plant species that were of interest for our goals, we collated traits related to dispersal, phenology and nitrogen fixation available from the TRY plant trait database (Kattge et al., 2011), as well as associations with mammals, birds, and herbivorous and pollinating insects from additional sources (see Table 1). The list of associated faunal species was refined to keep only species that occur in this habitat. Plant species frequency of occurrence values was used to rank plant species' relative abundance within the habitat type on a biogeographical scale, which we used to classify plant species dominance for a fixed species list for comparison with prioritized objectives (Table 1, see Table S1).
We then grouped the 163 plant attributes into nine broad categories based on the ability to support specific ecosystem functions or services (Table 1): bird trophic diet, bird herbivory, bird shelter, seed dispersal syndrome, Lepidoptera relationships (species specificpollination, herbivory), pollination syndrome, mammal herbivory, nitrogen fixation and flowering month. The range of attributes supported by plant species varied greatly, with some highly specialized plant species supporting only one attribute (e.g. Galium estebanii) while others support many attributes (e.g. Poa alpina, alpine meadow grass (56 attributes) and Sedum album, white stonecrop (58 attributes)) (see Table S1).
To assign attributes to species we used a binary classification scheme, where a value of 1 was used when an attribute was present in a given plant species, resulting in a plant species-attribute matrix.
In some instances, the presence of an attribute is dependent on the connection between the plant species and a species of other trophic groups, such as birds or butterflies (Lepidoptera). Some Lepidoptera species depend on different plant species at different life stages (larval herbivory vs. adult pollination/visitation), which we accounted for (Table 1). For birds, we connected trophic dependencies between attributes. For example, the plant species Aster alpinus (alpine aster) has a beetle pollination syndrome, and the bird Turdus torquatus (ring ouzel) feeds on beetles as part of its diet, so alpine aster is potentially an important habitat component for beetles, and the ring ouzel (see Table S1). Because we used categorical trait/attribute values here, we did not average traits across species when multiple values were found. If a plant species was represented in the data as having multiple dispersal syndromes, for example, we recorded them all. Additionally, we recognize not all plant species provide equal quantities of resources for other trophic levels and that our binary classification scheme currently does not reflect that. This approach is adaptable to represent different types of relationships flexibly; TA B L E 1 Plant species attributes, ecological role and data source of each attribute used, grouped by broad ecosystem function or service category. Complete list in Table S1 Broad attribute category (# of attributes) Plant attributes

Ecological significance and explanation of objective approach Source
Attributes related to trophic, dependencies, ecosystem functions and services Bird shelter (1) Shrub plant growth form Seed dispersal services (Cramp, 1978;del Hoyo et al., 2016) Herbivorous Bird Diet (20) 20 species of alpine bird that occur in these habitats (by expert opinion and available data). The herbivorous diet of each bird was identified, and connected with the plant species pool, each bird species is treated as an attribute Insectivorous Bird Diet (28) 28 species of alpine bird that occur in these habitats (by expert opinion and available data). The insectivorous diet of each bird was identified, and connected with the insectivorous pollination syndrome attribute of the plant species pool, each bird species is treated as an attribute  (Diaz et al., 2004;Fitter & Peat, 1994;Gachet, n.d.;Moretti & Legg, 2009;Poschlod et al., 2003) Mammal Herbivory (4) Ingested by mammals generally, and specifically herbivory by marmots, ibex, and chamois, key herbivores of this system Seed dispersal and grazing services (Andreoli et al., 2016;Bassano et al., 1996;Parrini et al., 2009) Nitrogen Fixation (1) Leguminous plant species Soil quality improvement  Flowering month (9) Represents every month February-October Provision of seasonal resources for pollinators (Aeschinmann, 2004;Plantarum, n.d.) Attributes used for comparison objectives Taxonomic Diversity/ Biodiversity by Genus (115) One plant species is selected from each taxonomic genera (115 genera total) within the defined regional species pool For this objective, a species from each taxonomic genus is selected randomly to represent a null representation of taxonomic richness (The Plant List, 2013) (Continues) for example, by weighting the importance of attributes among plant species. Here, we maintain a binary classification scheme of attributes to maintain comparability across objectives.

| Objective construction, comparison lists and prioritization
To construct objectives, we first set targets and create problems to be solved (Figures 1 and 2). We constructed five objectives  Frequency (1) Here, we use frequency of occurrence values obtained from the European Vegetation Archive to estimate which species are the most frequently occurring within these habitats on a European-wide scale, we then use this as a proxy for species 'Dominance' here. Frequency of occurrence values was used to rank species, and then the top n frequent species were selected to match the n of species required for the Comprehensive objective (Table 2) Dominance is associated with a contribution to carbon, nutrient and water cycling (Grime, 1998). Dominant species are often selected in plant species mixes for restoration, as they are known to represent key species of habitats or be helpful facilitators. We use this as another 'null' model for comparison with prioritized objectives Schaminée et al., 2016) TA  (5) Table 2).
To compare a species mix of dominant species to prioritized objectives, we sorted dominant species by frequency of occurrence values, and created a fixed list of 'dominant' species equal to the number of species in the Comprehensive solutions for direct comparison between the performance of the two species lists in terms of representing attributes that potentially support particular ecosystem functions and services. We consider a single presence of an attribute so that objectives and comparisons could be directly contrasted.
To efficiently find the smallest number of plant species that met the target-based objectives (Table 1), we used the 'minimumset' problem formulation which is commonly applied to spatially explicit decision-making that cost-efficiently meet targets for conservation features (e.g. habitats, species ranges or ecological processes; Possingham et al., 2000). We adapted inputs to apply it to our non-spatial problem (Hill et al., 2014; see Appendix S1). To  solutions the performance relative to the optimal (gap = 0). For all random solutions for comparison, we used 'add_shuffle_portfolio' (instead of the gap portfolio). This randomly reordered data prior to solving problems, so plant species were selected under different data formulations to produce a random selection process. These problems can also be solved using MARXAN, and we provide more details on using these applications for these applications in the Supplementary Information (see Appendix S1).

| Evaluation
In comparing and evaluating approaches to plant species selections for species mixes, prioritizr produces two important outputs: optimal solutions to meet targets for objectives (in this case, a plant species list); the feature representation indicating the number of plant attributes represented by a solution, relatively (to possible maximums) or absolutely (total number). We used the first optimal solution of each objective to compare the attribute sum of the plant species selected (see Table S1). We compared the mean values of the attribute sum using a Kruskal-Wallis chi-square test to assess differences in the total number of attributes (attribute sum  Table S1) and calculated the percentage of attributes captured compared to the total number of attributes possible for selection in each broad attribute category (Table 1).
We compared our results to an ad-hoc selection of species that we approximated using a random species selection process. We

| RE SULTS
Across the five ecological objectives, the number of plant species needed to meet each objective's targets varied widely. For example, the targets for the 'Bird' objective were met with only five plant species, targets for the 'Pairwise Lepidoptera + Plant Rich Genus' objective required 119 plant species (Figure 3). We also found high variability in the number of attributes (the attribute sum) captured by the individual plant species selected in the solutions (max:59; min:2; see Table S1). The plant species selected in the 'Comprehensive' and the 'Lepidoptera Relationship' objective had a significantly higher attribute sum overall (Kruskal-Wallis s 2 = 146.68, p = <0.001, Table S3, Figure 4, see Figure S1) than found in the other objectives.
This prioritization approach favours plant species with a high attribute sum, yet also prioritizes plant species that supports unique or rare attributes, as these species may be considered irreplaceable ( Figure 4). In the case of the Comprehensive and the Lepidoptera Relationship objective, many plant species were irreplaceable, representing specialist relationships. In contrast, in the bird objective, many plant species were of variable importance, and thus could be interchangeable ( Figure 4; Table S4). Objectives that did not set out to prioritize a specific category of ecosystem function had variable performance. For example, the 'Bird' objective performed poorly for all plant-pollinator-related ecosystem functions, capturing <25% of the attributes needed to support plant-pollinator, nectar and larval functions ( Figure 5). This is unsurprising given the 'Bird' objective only needed five plant species to achieve the targets. Alternatively, the 'Comprehensive' objective, which aimed to represent each of the 163 attributes found across the entire plant species pool once (and did so with 37 plant species), met 100% of targets set (1 of every attribute). However, the species prioritized in this objective performed no better than the random species selection for representing Genus and Family levels of plant taxonomic diversity representation (Figure 4).
Overall, the random selection of plant species performed well for ecosystem functions that are supported by common attributes across plant species (e.g. bird trophic, bird herb, bird shelter), but worse than our prioritization when the ecosystem function is supported by a highly specialized attribute (e.g. plant-pollinator relationships; see Figure S2). In general, the smaller the number of randomly selected plant species, the worse the performance for providing ecosystem functions and services. Even when large numbers of randomly selected plant species are considered, provision of some trophic relationships or ecosystem function and service groups was found to be low (see Figure S2). and services can in some cases be unsupported and low while plant species richness is high. We confirm that a random draw of plant species from the regional plant species pool cannot be assumed to support other trophic groups and ecosystem functions and services.

| D ISCUSS I ON
This has important implications for the design and implementation of species mixes for restoration projects which aim to reach multiple restoration objectives such as plant diversity, higher trophic levels and certain ecosystem functions and services tied to plant species identities.  (Knight et al., 2005), and in restoration, there is the opportunity to directly support these connections between organisms positively facilitate regeneration F I G U R E 5 Trade-offs between number of plant species selected for each objective, and the corresponding proportion of ecosystem function targets captured compared across objectives. Optimizations were run 100 times for each targeted coloured objective and all unique runs are shown here as points. Grey points represent the mean, upper (75%) and lower (25%) quantiles of runs for randomly selected species in intervals of 5 from 5 to 120 plant species processes and ecological networks (Harvey et al., 2017;Valiente-Banuet et al., 2015) through this framework. When considered this way, one can ask if the species pool used in restoration is providing adequately for the species pool of other trophic levels within that habitat while balancing multiple targeted outcomes.

| Prioritizing functional attributes
Conversely, depending on how plant taxonomic diversity is defined (representing one species from every taxonomic Family or Genus), it is not always represented well by objectives prioritized for attributes, or by randomly selecting species, but can be captured efficiently through targeted selection. Additionally, both attributes and plant biodiversity can be captured efficiently together when both are set as targets (Pairwise objectives). Seed mixes matter for restoration success and can be optimized according to many factors (Barr et al., 2016), but require the balancing of multiple targets which is a complex multivariate decision-making task that can make use of decision-support tools as demonstrated here.
Additionally, seeding and planting treatments for restoration are restricted by many confounding constraints including budgets, labour and project size and so restoration species mix treatments are often quite low (Barr et al., 2016). Where constraints are present, prioritizing plant species to optimize particular targets can be a potentially beneficial method to decide which species to include in low-diversity treatments. This method has similarities to methods for filtering plant species lists based on particular targets (Brudvig & Mabry, 2008). This method also has similarities to selecting plant species to optimize functional diversity, or for the convergence of targeted plant trait representations for particular environments (e.g.
traits that increase chances of survival in dry, harsh environments; Laughlin et al., 2018;Wang et al., 2020Wang et al., , 2021. However, our approach presented here is distinct from these other approaches, and it offers the unique advantage of optimizing targets according to flexible constraints and offers quantitative support for comparing different options easily both in terms of targets and cost. There is surely opportunity for future work to combine these approaches in the future.

| Indications and further development
To test this proof-of-concept, it was necessary to make some simplifying assumptions. Focusing on a study system with relatively good knowledge on frequently occurring plant species, we selected attributes based on available data in the target system. Rare plant species, which have not been thoroughly studied, are often documented as having few attributes, resulting selection bias towards representing common species. A prioritized solution can only be as good as the data available, and the prioritization objectives set out here are limited by the available data. Generalized data on pollination syndromes or seed dispersal syndromes of plant species can be limiting, as these relationships can be habitat specific. Although we used the best data available from a trait database and field guides, we recognize that next steps should include an improvement on data used. These data include plant-insect associations for additional insect taxa (e.g. wild bees as pollinators or plant and leafhoppers as herbivores), and of improved occurrence data (for our work, no data on the altitudinal occurrence of moths were available). Local entomological specialists can help to compile realistic lists of plant-insect interactions, and we postulate that this method could also make excellent applied use of pollinator network or food web data across trophic levels.
A 'Comprehensive' species mix (Figure 1, Comprehensive objective) is one that captures all desired attributes within the plant species selected. Here, we focused on comparing the comprehensive approach to others to answer our questions and to demonstrate this proof-of-concept. In our approach, targets were set to ensure a minimum of one attribute was present in solutions, allowing for a direct comparison between objectives, but this means other plant species were categorized as redundant when not selected. In practice, including several species with the same attributes may be desirable as an insurance plan for species that do not germinate or that exhibit intraspecific variation. By including some redundancy of attributes, one can minimize the chance that this attribute will not be represented in the realized restored community. Three types of abundance are relevant for reaching restoration objectives: the species mixes used, including quantity and quality of each species Shaw et al., 2020), the final abundance of each species in a restored plant community and the final expression of each species attribute (including interspecific variation). Our approach can be adapted to consider abundances of attributes at all three levels, which may be a subject of further research.
Similarly, we assume that the cost of including a plant species is the same to test our questions independent of costs, but the approach can and should account for cost variation to acquire, store and reproduce seeds as this will likely hold great influence on prioritized solutions in practice (Jiménez-Alfaro et al., 2020). Reporting on the costs of conservation and management actions is largely inadequate and non-standardized (Iacona et al., 2018), and we know from previous research that only a small proportion of seeds are usually available for purchase (Ladouceur et al., 2018), and must be collected by hand, which is costly . Here, while we use the 'minimum-set problem', objectives can also be created that meet targets best within cost constraints. Prioritization approaches could also guide future efforts for native seed supply and policy by informing collection, farming and storage for an expanded restoration species pool. Furthermore, including these real costs in decision-making frameworks can help to plan efficient projects, and can be used to communicate budget constraints and needs in a robust way.

| CON CLUS IONS
This proof-of-concept is the first step towards framing future empirical research in optimizing species mixes for the ecological restoration of natural ecosystems. We call for empirical field tests for this approach to take place, which will require bringing together interdisciplinary collaboration across subfields of ecology and conservation. We provide a transparent and robust approach that could move restoration efforts towards prioritizing plant species to maximize targets and minimize costs offering quantitative decision-making support. This approach could be applied to any system and/or targets which could also contributed to many stages of restoration decision-making and could play an important role in delivering efficient, targeted solutions. However, similar approaches will need robust ecological data to be applied to specific cases studies and restoration targets, preferably at regional or local scales.