How Do Multi-criteria Assessments Address Landscape-level Problems? A Review of Studies and Practices

Viewing the landscape as a spatialized social-ecological system allows identification of specific management challenges: integration of multiple views, multiple levels of organization, complex spatial-temporal patterns and uncertainties. Multi-criteria assessments (MCAs), which allow the comparison of alternative actions when multiple interests collide, are considered adequate to support landscape management. However, there is no consensus about how they should be applied and can integrate both multiple views and spatial dimension. We conducted an extensive quantitative and qualitative literature review targeting MCAs with a participatory and spatial approach. Our results suggest that (1) for sustainability assessments, participatory and spatial approaches endorse different rationales and hybrid methods are not so common; (2) within those methods, only scenario-selection methods (as opposed to design methods) can integrate spatially-explicit, spatially-implicit, place-specific, and overall values; and (3) current applications, which aggregate values ignoring their spatial and social distribution, do not coincide with the nature of landscape-management challenges. In addition, they give little importance to the structuration of information and to collective deliberation. We conclude that, in the absence of a good match between spatiality and participation, MCAs should, for now, be handled as insightful but distorted tools to explore and structure landscape-level management problems.


INTRODUCTION
Landscapes are complex social-ecological systems (SES) because many human and natural processes mutually interact (Bastian, 2001;Cumming et al., 2012;Naveh, 2000;Wu, 2006) and because they are shaped by a social history (Antrop, 2000;Pedroli et al., 2006), which promotes perceptions, values or expectations that differ spatially and among individuals. This makes it challenging to collectively define a desirable future for a given landscape. Many possible ways exist to address this issue; one of them is multi-criteria assessment (MCA). Nonetheless, there is no consensus on how to apply this method given the nature of landscape-level challenges.
A first challenge comes from the complexity of processes and multiple interactions. Because landscapes are characterized by interdependencies between human societies and their environment that originate from a coevolutionary history (Berkes et al., 2000;Costanza et al., 1998;Daily, 1997;Kallis and Norgaard, 2010), they exhibit non-linear and cascading effects that make their trajectories of change impossible to predict (Kinzig et al., 2006;Walker et al., 2004). It is generally acknowledged that command-and-control approaches Holling and Meffe, 1996) and risk assessment (Linkov et al., 2014) are ill-suited in this context. Instead, scholars advocate exploring possibilities for change (Berkes et al., 2002;Olsson et al., 2006) and adopting an adaptive management approach (Linkov et al., 2006;Plummer, 2009). Another challenge with landscape-level problems arises because the different groups concerned hold different and sometimes irreconcilable values (Gómez-Sal et al., 2003;Hunziker et al., 2008;Swedeen, 2006). This problem of multiple views, which is common to many decision-making situations, has led to recommendations to include non-experts in evaluations of the quality of decisions (Funtowicz and Ravetz, 1990) and explicitly consider incommensurabilities 1 (Martinez-Alier et al., 1998;Munda, 2004). Deliberative approaches and MCAs involving multiple stakeholders are considered particularly well-suited to operationalize these principles (Frame and Brown, 2008;Munda, 2004).
1 Specific concepts are defined in the appendix.

MATERIALS AND METHODS
This literature review follows three steps (the overall method is described in Fig. 1). First, we performed a lexicometric analysis of a large sample of studies to characterize the position of sustainability assessments, participatory approaches and spatial approaches within the wide spectrum of multi-criteria methods (Section 2.2). Second, we qualitatively classified applications of multi-criteria approaches mixed with participatory and spatialization methods in the field of natural resource management (Section 2.3). Our aim was to provide a typology of existing methods with their general steps. Third, we focused on a specific type of methods arising from the typology, "scenario-selection support method", and undertook a detailed qualitative analysis of the corresponding case-studies (Section 2.4).
We generated three datasets of studies of decreasing size using the Web of Science database, corresponding to the three steps of our review. For better traceability, we summarized this selection process in a PRISMA diagram 2 (Moher et al., 2009) ( Fig. 1). A large dataset of abstracts (10,691) was selected to analyze recent trends (2005)(2006)(2007)(2008)(2009)(2010)(2011)(2012)(2013)(2014)(2015) in sustainability assessments within multi-criteria methods. For the two subsequent steps, the time span was extended to all available years  to embrace a wider diversity of research; nonetheless, few records were published before 2005.
The three steps reflected in the PRISMA diagram are detailed in the three following subsections.
Lexicometric analysis compresses complex information contained in large numbers of texts.
Reorganizing and classifying segments of texts and providing textual statistics promotes understanding of the meaning and context of use of utterances (Lebart et al., 1998). We used this method to identify trends in the current use of multi-criteria methods throughout the observation of lexical similarities among abstracts.
We collected information that characterizes the heterogeneity of research referring to multicriteria methods. We compiled the 10,691 abstracts we selected into a single textual corpus and used IRaMuTeQ software (Ratinaud and Marchand, 2012) to implement Reinert's method (1993)  showing the presence or absence of words (except pronouns, conjunctions, and some adjectives) in a given abstract. IRaMuTeQ clusters abstracts by performing descending hierarchical classification iteratively: abstracts are divided into two significantly different groups (p-value < 0.05) according to the presence or absence of words, and then the cluster with the most abstracts is divided into two new significantly different groups, etc. Clustering stops when the number of clusters predefined by the analyst is reached or when the largest cluster contains no significantly different vocabulary. A chi-square test is then calculated for the classification of each word to assess the significance of the association between a word and a semantic class. This enabled us to obtain a broad view of multi-criteria methods and then identify subgroups relevant for analysis.
We also assessed the classification via correspondence analysis, in which variability in the vocabulary and its distribution among clusters (assessed as a chi-square value) is statistically explained by inertia. For each inertia factor, correspondence analysis provides a statistical basis for interpreting why certain clusters are similar to or different from others.

1.3.
Step 2. Qualitative classification of multi-criteria assessments that use participatory and spatial approaches.
In the second step of our analysis, we emphasize the participatory and spatial approaches of MCA.
We collected research that links these three aspects (i.e., multi-criteria, spatialization, participation) through a database search and manual screening. "Spatialization" meant that alternatives and/or the assessment results had to be spatialized; "participation" meant that people other than the authors had to be involved at some stage of the research. We also removed all reviews, theoretical developments and studies that introduced a multi-criteria tool without applying it (even though they could provide interesting insights for interpretation).  Among the 222 studies collected, 126 met those eligibility conditions, and among them 74 belonged to the field of natural resource management. The others referred to waste-or pollution-management issues, urban or infrastructure development and health.
We systematically read the abstracts of the 74 studies dealing with natural resource management, with a special focus on their objectives and methods; when necessary, we also read the studies' methods sections. We focused more on the stages in which the inclusion of a spatial or participatory component to MCA adds to or modifies the way the method is developed. We summarized the most frequent method patterns according to their objective.

Step 3. Systematic qualitative analysis of case studies for scenario exploration.
To address our third objective, we analyzed a subset of the previous studies. We only investigated articles that developed a method for exploring landscape management alternatives and possibly selecting one of them. We ended up with 10 studies, some of them covered by more than one article. When necessary, we used additional articles referring to the same research project and contacted the corresponding author to clarify certain points or to obtain additional information.
Two types of characteristics were examined for each study. First, general characteristics of MCA were analyzed: the problem structure (i.e. alternatives, criteria, stakeholders and their judgments) and the decision analysis (the weighting system and the aggregation procedure used to produce the final evaluation of alternatives) (Kiker et al., 2005;Malczewski, 1999).
Second, we investigated characteristics related to the landscape viewed as a spatialized SES, focusing on participation of multiple stakeholders (i), representation of multiple levels in the assessment(ii), consideration of spatial and temporal patterns and relationships (iii), and management of uncertainties (iv).

Main clusters of multi-criteria methods: the vocabulary of sustainability assessments
In the abstract corpus (10,691 records), we performed a three-cluster classification, which allowed 9,450 abstracts (88%) to be classified (Fig. 2). The most significant forms of words, tool-words excluded, define three distinct "lexical worlds".
We identified a first cluster of abstracts (3,189 records, 34% of the total) with the most significant forms relating to: integrating economic, environmental, social and institutional issues ("environmental", "economic", "social", "policy", "public") considering future consequences and uncertainties ("impact", "assessment", "options", "management", "potential") engaging the public ("stakeholder", "public") Gasparatos et al. (2008) depicted these different aspects as shared properties of sustainability assessments and added equity considerations (intra-and inter-generational) to this list. Equity considerations, however, are not directly reflected by the most significant words in the "sustainability assessment" cluster. The first word in this semantic field, "equity", appears in 312 th position (still significant for this cluster) because it occurs rarely. We conclude that social justice issues are, in most cases, absent from sustainability assessments based on a multi-criteria method. Sustainability remains mainly understood as the fulfilment of competing goals rather than a trajectory of change that reconfigures social interactions.
A second cluster of abstracts focuses on optimization problems ("optimization", "Pareto", "solution", "design", etc.) aiming at designing Pareto-optimal 3 solutions for a limited set of objectives and constraints (Linkov et al., 2006). Decision-analysis procedures ("decision", "rank", "selection", "aggregation") are grouped together in the last cluster of abstracts (cluster 3). Considering that no optimal solution can be found, a variety of procedures exist to rank/sort alternatives (Guitouni and Martel, 1998), each of them corresponding to a specific definition of what makes the best compromise (the closest from an ideal point, the one that is not outranked by others, etc.).

Rationales underlying MCA methods: the specificity of spatial approaches
3 Resource allocation is considered optimal (sensu Pareto) when an increase in one objective cannot be achieved without worsening another objective. The total number of situations that satisfy the Pareto criterion forms a boundary limit from which no improvement can be made given the available resources. The existence of such a boundary assumes that objectives are perfectly substitutable. We generated an additional classification to see how the sustainability assessment cluster would split up, adding two more successive phases of clustering. The abstract corpus was consequently divided into 5 classes (8,964 abstracts included in the new classification, ~84% of the corpus). We were then able to distinguish two new subgroups of abstracts: strategic business decision-making (within the "decision analysis" cluster, former cluster 3) and spatial approaches (within the "sustainability assessment" class), the latter being of special interest for our purpose. The "optimization problem" cluster, (former cluster 2) remained unaffected by this new classification. Projection on the two first axes of the correspondence analysis ( Fig. 3), which summarize 69% of the total inertia, reveals that spatial assessments constitute both a marginal (approx. 1/4 of the abstracts on sustainability assessments) and specific branch of sustainability-assessment approaches. We interpret the two first factors as reflecting what is understood as a rational decision (axis 2, 29% of total variance) and what makes a scientifically-sound decision (axis 1, 40% of total variance).
The first axis differentiates multi-criteria methods that rely on grounded, contextualized information from those that mainly rely on mathematical equations to find a solution. By "grounded", we mean that theory proceeds from the accumulation and analysis of place-based data. Since they seek to capture multiple indicators and multiple views, sustainability assessments imply such bottom-up view of science, providing they not only try to formulate ad-hoc solutions but also create new knowledge. The opposite pole of axis 1 represents another view of how science interacts with decision-making, in which there are "rules of nature" that can be formalized (by scientists) and then applied to any situation (e.g. to support decision-making). According to this view, generic models and algorithms are the elements that validate the results obtained. Decision analysis, which focuses on finding the best compromise when no optimum solution exists, and optimization methods, which focus on finding solutions that maximize utility under a set of constraints, are two different methods that share the assumption that mathematical formalization is the best way to reach a valid solution.
Axis 2 reflects the opposition between substantive rationality and procedural rationality behind multi-criteria methods (Simon, 1976). Under a substantive rationality hypothesis, rational behaviors allow the fulfillment of a set of goals under a set of constraints (from the external environment, its perception, or internal characteristics of the decision maker). This realm assumes a "heroic" (Béjean et al., 1999) decision-maker who, according to Simon (1990: 195): "contemplates, in one comprehensive view, everything that lies before him. […] He has reconciled or balanced all his conflicting partial values and synthesized them into a single utility function that orders, by his preference for them, all these future states of the world". This idea is clearly embedded in optimization procedures. However, the substantive rationality paradigm is also dominant in spatial assessments. The use of spatial data is thought to "empower the decision-maker" (Densham, 1991) by providing both more precise and more usable information. The environment (objective and subjective) is considered fixed and the coexistence of different value schemes and conflicting preferences not allowed.
In the realm of procedural rationality, behaviors are rational because they come from an appropriate individual deliberative process in the face of incomplete information, an unknown future, or shifting beliefs and values (Simon, 1976). Studies related to decision analysis, since they model uncertainty and the interaction of multiple views towards a compromise, typically assume a procedural rationality (extended from the individual to the group) (Garmendia and Gamboa, 2012;Munda, 2004). Sustainability assessments following a participatory approach ( Fig. 3, upper-left corner) also fall in this category because deliberation is the pivotal process for integrating heterogeneous information and values (Vatn, 2005). Nonetheless, these approaches often fail to accommodate the evolutionary nature of preferences and the existence of non-utilitarian values (values that cannot be reflected in competing individual interests). Rather, they assume that one's motivation can be captured in a unique order of preferences (Béjean et al., 1999). They therefore only partially fall under procedural rationality paradigm.
We have just demonstrated how two branches of sustainability assessment distinguish themselves. Though common in their place-based anchorage, sustainability assessments following a spatial approach and those following a participatory approach endorse fundamentally different philosophies about what constitutes a good decision (the process or the availability of information) and the context under which a decision is to be taken (irreducible uncertainties and conflicting values or a well-defined decision space). Nonetheless, bridges between these two branches exist and are crucial if one wants to address multi-stakeholder landscape-level issues. We examine these bridges in the following section.

How MCA combines with participatory and spatial approaches. Synthesis of methods and objectives.
In this section, we present results of our qualitative analysis that classified studies using participatory and spatial approaches (Fig. 4). We distinguish three types of objectives, hence three types of methods: scenario-design support, scenario-selection support, and spatial assessment of real situations. The categories of both scenario-design support and scenarioselection support methods are future-oriented, leading to management recommendations.
Method steps can be shared between real-world and future-oriented methods even when their objectives differ (prescriptive as opposed to descriptive). We found two studies (3 records) combining scenario-design support and scenario-selection support objectives (detailed in section 3.3.2).
Regarding the two families of future-oriented methods, their differences in objectives are reflected in different method steps (Fig. 5). Scenario-design support methods deal with a location or spatial allocation problem (van Herwijnen and Janssen, 2007): Which areas should be managed or preserved first (prioritization)? How best to allocate land-uses in the landscape (zoning/ mapping proposals)? Those questions refer respectively to multi-attribute or multiobjective problems. In the first case, every location is scored according to a set of attributes (van Herwijnen and Janssen, 2007). Then, value maps are overlaid and aggregated, generally with a weighted linear combination procedure (Malczewski, 2000), to define the best area(s) or rank sites for a given purpose. This method is also used for suitability assessments. In the second case, a set of goals are defined and multi-objective optimization algorithms produce land-use plans or spatially-explicit management proposals. Conversely, scenario-selection support methods differentiate and often rank management alternatives based on their effects (impact assessment) and sometimes their implementation (generally cost, but sometimes applicability or flexibility; e.g. Sahin and Mohamed, 2013). A scenario can be evaluated and possibly selected according to both spatially-explicit and nonspatial criteria. An aggregation procedure can be performed first on criteria (producing a landscape-level score for each criteria) or on alternatives (producing total performance maps) (see van Herwijnen,1999, cited by van Herwijnen andJanssen, 2007), for different possible pathways). The involvement of non-academic participants, including local experts, can occur at different stages. In such cases, identified criteria and/or weighting are designed to reflect stakeholder preferences. The issues of how to select participants, how to elicit their preferences and how to combine them are the same for non-spatial evaluation (see e.g. Garmendia and Gamboa, 2012;Hajkowicz and Collins, 2006;Reichert et al., 2015). Also, some criteria can be evaluated on a participatory basis when they integrate a sensible dimension (e.g. visual aspect (Sheppard and Meitner, 2005)); risk perception (Raaijmakers et al., 2008)); place attachment (Newton et al., 2012;Nordström et al., 2011) or when the evaluation of performances relies on the statements of local experts (e.g. Arciniegas and Janssen, 2012).
Through this clarification exercise, we highlighted consistent patterns in scenario-design and scenario-selection methods. The way stakeholders are involved does not appear to differentiate methods. Stages with participation -and the extent of that participation -reflect the diversity of scientific postures in action-research, which is independent from the operational objective pursued. In contrast, the position of the MCA within the decisionmaking process and the nature of criteria and values do distinguish scenario-design support from scenario-selection support. In the former, alternatives are generally not defined a priori, and the decision-making process is in its exploratory stage; also, criteria and values are necessarily spatially-explicit. In the latter, alternatives are established a priori. They can be realistic plans identified by managers (when the decision process is in an advanced stage) or contrasting images of the future created to support learning. With scenario-selection support methods, alternatives can be spatially-explicit or not (e.g. a normative change) and the assessment of alternatives can accommodate both spatially-explicit and non-spatial values.
Assessment is therefore more comprehensive, which suits a holistic and integrative view of the landscape. Because of these possibilities, we focus on the potential of scenario-selection methods to compare landscape-management alternatives.

Overview of selected studies
In our corpus, we found 11 studies (12 records) that implement a multi-criteria scenarioselection method at the landscape level, i.e. studies that use, at some stage, the input of nonacademics to compare a set of landscape-management scenarios to highlight those best-suited (Table 1). The study of Raaijmakers et al. (2008) was excluded as there was no integration between the stakeholder assessment of alternatives (not spatial) and the simulation of their impacts using a GIS-based model. The 10 remaining studies deal with land-use planning (4), ecosystem conservation (3), water management (2) and forest management (2). Our sample covers the gradient of research postures that characterizes integrated assessments van Asselt and Rijkens-Klomp, 2002). Some of the studies we analyzed hence implemented a "lighter" participatory approach, corresponding to a situation of consultation or information . In these cases, stakeholders are only present at the very beginning of the process (associated with generating alternatives) or merely exist as assumed recipients/users of assessment results. They can also be absent but different values nonetheless considered through archetypical weighting schemes (in the case of Linhoss et al., 2013).
We identified two studies that combine scenario-selection and scenario-design approaches, and hence promote iteration for adaptive management. One (Nordström et al., 2011(Nordström et al., , 2010 used a scenario-design approach to generate and select three realistic but contrasting scenarios of forest management, which were consequently assessed. The other (Arciniegas et al., 2011;Arciniegas and Janssen, 2012) followed the oppose approach: scenario assessment was used as an exploratory phase prior to generating land-use plans. Although they are rarely used jointly, scenario-design and scenario-selection methods are not mutually exclusive. Table 2 summarizes the options adopted in the case studies reviewed.  (Choo et al., 1999;Garmendia and Gamboa, 2012;Guitouni and Martel, 1998;Munda, 2004;Vatn, 2009).

Results of qualitative analysis
In the 10 studies we reviewed, alternatives and criteria are either developed by researchers or by a set of stakeholders. Specific reference frameworks (for ecosystem services, ecological functions or management standards) are sometimes used. In some cases, definition of alternatives is a central part of the assessment method and results from a formalized exploration-selection-specification process (in Fürst et al., 2013;Nordström et al., 2011).
Often, but not always, criteria are translated into indicators (in other cases, this hierarchical structure is absent); definition of these indicators is managed by researchers and experts, without feedback from other participants.
The process for selecting the stakeholders included in the assessment is often not formally addressed. Two studies (Nordström et al., 2010;Sheppard and Meitner, 2005), which defined their study area as conflict-ridden (both around the issue of forest management), used stakeholder analysis to identify stakeholders and group them according to their assumed interests. derive weights was found in an expert workshop in Arciniegas and Janssen (2012). Generally, stakeholders are asked to score criteria according to their importance. In one study (Nordström et al., 2010), stakeholders were asked to state preferences using a pairwise comparison of criteria.
More originally, Newton et al. (2012) and Linhoss et al. (2013)  In the spatially-explicit assessments we analyzed, basic units or compartments are defined according to their homogeneous functioning (Chakroun et al., 2015;Janssen et al., 2005;Manoli et al., 2005) or their use (Arciniegas and Janssen, 2012). Another option is to use a cell grid (Fürst et al., 2009;Newton et al., 2012). of spatial aggregation for two ecosystem services using landscape metrics 4 . Limits of spatial aggregation can also be overcome through an overall evaluation: Sheppard and Meitner (2005) attempted to do so by asking stakeholders to directly assess semi-realistic landscape visualizations of competing alternatives.
Temporal patterns are conspicuous by their absence, with the notable exception of Manoli et al. (2005), who first spatially aggregated each criterion considered (e.g., cost, water availability) and then considered second-level criteria describing temporal behavior.
(v) Uncertainties Finally, uncertainties are generally not considered. When they are, the most common practice is to perform a sensitivity analysis on the weighting systems, which reflect uncertainties in the preferences for each criteria (Chakroun et al., 2015;Janssen et al., 2005;Newton et al., 2012). Linhoss et al. (2013) chose to use of a stochastic multi-criteria decision analysis in order to test a case of uninformed weights (the whole weighting space is explored and a rank acceptability index is calculated, describing the percentage of times one alternative ranks as "most-preferred"). Uncertainties about future changes are addressed in two studies (Linhoss et al., 2013;Manoli et al., 2005), in which climate is considered the main driver of change.
Uncertainties in judgment are addressed in the two studies that attempted to structure stakeholder participation in a conflict-ridden context (Nordström et al., 2010;Sheppard and Meitner, 2005) and in the study of Linhoss et al. (2013) that uses a stochastic procedure.
Another way to handle uncertainties is to design a flexible decision-support tool that allows modifications. Some tools are also user-friendly (e.g. the touchtable of Arciniegas and Janssen (2009); the "Pimp your landscape" software of Fürst et al. (2009)). It is not clear, however, whether these tools can be modified by end-users.
4 Landscape metrics refer to measures or indices that consider the distribution of patches (or groups of patches) within a landscape mosaic. They can characterize either a landscape's composition (e.g., abundance of forested patches, evenness of different patch types) or spatial configuration (e.g., measures of patch density, shape, or connectivity).

Critical discussion of observed practices.
This systematic qualitative analysis enables us to identify and discuss several research gaps in the application of MCA methods to assess landscape management alternatives.
(i) Lack of attention to problem structuring Problems observed in MCA in general are also found in our analysis. Among them, the lack of attention to problem structuring, which is crucial for establishing a shared platform for deliberation and effectively linking assessment to decision making, is striking (Boroushaki and Malczewski, 2010;Giampietro, 2003;Hajkowicz and Collins, 2006;Ramsey, 2009). This loophole is particularly prevalent in the generation of alternatives, selection of stakeholders, and definition of indicators.
The selection of stakeholders, alternatives, criteria and indicators constitute a "compression" of the information space (Giampietro, 2003) that conditions the quality of the MCA outcome. As shown by this literature review, alternatives and criteria are often defined with relatively little involvement of local stakeholders, although the conditions of their involvement are often not fully clarified or formalized. Institutional analysis -at worst to identify the diversity of interests and "legitimate" people to participate in the evaluation and at best to gain knowledge about their power relations -is a noteworthy method to justify the choices made about participants, scenarios and criteria (De Marchi et al., 2000;Munda, 2004;Xenarios and Tziritis, 2007). Also, if alternatives are used to expand the decision-space -i.e. alternatives represent contrasting images that stimulate stakeholders' creativitythen formalized procedures (computer-assisted or not) that enable moving out of the classic set of "sketched" alternatives can be useful (Groot and Rossing, 2011).
Defining indicators related to criteria is commonly considered the task of researchers and experts, which is consistent with a positivist view of science. However, using MCA, especially in conjunction with participatory methods, should force researchers to endorse a post-normal paradigm for evaluating sustainability (Funtowicz and Ravetz, 1994). way of organizing information (i.e. a specific view of the problem), then it modifies the status of indicators in MCA: indicators are not only "proxies" that provide a measurement of a criterion's performance but are also objects for collective deliberation (Frame and O'Connor, 2011).
(ii) Mismatches in the use of aggregation procedures Another recurrent problem with current practices is that common aggregation procedures (i.e. compensatory methods) are not consistent with the way value judgments are made or with sustainability issues characterized by incommensurabilities. The first problem that was repeatedly observed lies in the meaning of weights, which relates to the type of question stakeholders are asked (Choo et al., 1999). In most studies that integrate stakeholder preferences, stakeholders are asked to allocate points between criteria and then to score or rank them. Compensability is generally not addressed as such because scientists do not ask "By how many times is this more important?" or "How many units of one criteria would compensate for losing one unit of another?", etc. (Choo et al., 1999;Garmendia and Gamboa, 2012). Nonetheless, compensatory methods are usually used to aggregate preferences. Participants could learn about the conceptual bases that distinguish, for instance, compensatory aggregation from outranking methods through the questions they answer. Therefore, defining a "good" question to elicit preferences could improve the general challenge of making the assessment process more transparent for and controllable by those involved in decision-making (Bell et al., 2001).
A second inconsistency lies in the idea that compensatory aggregation methods suit sustainability assessments. If we consider the question of sustainability under a "non-substitutability" hypothesis (strong sustainability, sensu Neumayer, 2003), then values of one criteria cannot compensate values of another criteria, which excludes compensatory aggregation. Garmendia and Gamboa (2012) and Munda (2004) suggest using outranking methods to avoid this bias. Ahrenz and Kantelhardt (2009) tested one outranking method (Electre-II) and concluded that the transparency of the method is limited, so that the choice between compensatory and non-compensatory methods for landscape planning decisions should account for the context of the assessment. Likewise, the ordered weighted averaging aggregation method allows partial compensability among criteria. In a GIS environment, it has been used mainly for scenario-design support (e.g. prioritizing areas for forest conservation, Averna Valente and Vettorazzi (2008) adjusting forest-management plans, Greene et al. (2010).
Nonetheless, its mathematical and computational sophistication (Aliyu and Ludin, 2015) seriously limits its use in participatory settings. This trade-off between transparency and integration of complexity should therefore be clearly addressed and discussed by scientists to justify their aggregation choices.
(iii) Lack of consideration of the potential of GIS to enhance deliberation In sustainability assessments, MCA methods often aim to produce group or societal solutions (Boroushaki and Malczewski, 2010;Garmendia and Gamboa, 2012). In the studies selected, the "group" solutions achieved are intermediate positions between the preferences of the various stakeholders (e.g. a"compromise", sensu Vodoz, 1994). Preferences are considered immutable throughout the evaluation process, which can, at best, enhance the mutual understanding of participants. Only one of our 10 studies effectively links spatial evaluation to deliberation (Arciniegas and Janssen, 2012). In this study, maps not only provide a spatialized representation of performances, but are also intermediate objects for exchanging views (for analysis and negotiation). In the other studies, social learning remains understood as a "by-product" and not as a core process fundamentally attached to the evaluation activity.
We support the rehabilitation of deliberation in spatially-explicit evaluations, understood as an opportunity to dialogue, understand the position of others and eventually produce novel solutions underlying a change of view. Such a deliberative perspective reconnects spatial assessments with the philosophy of collaborative GIS (Boroushaki and Malczewski, 2010). Conditions to foster collective deliberation encompass: establishing inclusive platforms for exchanging views (considering the dominated actors and the silent mass of future generations), tracing the history of changes in the evaluation matrix, engaging reiterations, and analyzing the reasons that led to a group solution or impeded its emergence. This methodological posture suggests that the coupling of GIS tools with MCA should turn toward the former's potential to stimulate exploration, understanding and redefinition of the decision problem by those who are involved in the evaluation (Malczewski, 2006;Ramsey, 2009  (iv) Forming stakeholder groups based on an a priori affiliation Delineation of the stakeholder groups in the evaluation is crucial to allow for comparison of contrasting views. This task differs from the identification of knowledge-holders (to define indicators) and stakeholders (to choose diverse criteria and elicit judgments about them). The delineation of groups for comparing alternatives comes later, as a way to organize the diverse opinions previously explored and to identify possible alliances and conflicts. The common practice for forming groups that we observed (when groups are considered) is reliance on social and/or professional affiliation (e.g., farmers, environmentalists, water managers). However, it supposes that social and/or professional affiliation serves as a good proxy for value systems, i.e. value systems are considered homogenous within each social or professional group.
Leaning on other research works, we suggest alternative guidelines to form groups of participants and reflect the diversity of competing value systems. The first option is strategic: it consists in analyzing stakeholders' interests and relationships to identify "key" participants, e.g. participants who have a legitimate influence on the problem (e.g. De Marchi et al., 2000). In a landscape context, this option also supposes to make clear what scale of interest drives one's value judgments (Zia et al., 2011). A second option is to construct preferences archetypes as done in Ravier et al. (2015) from the statistical clustering of individual preference trees. The various preferences archetypes reflect the diversity of value systems without a priori assuming that these archetypes would coincide with socially instituted categorization. We believe this method is useful in conflict-prone situations, which are often encountered at the landscape level. A third option is to let participants define the group to which they belong, as done by Ahrenz and Kantelhardt (2009). This way, part of the structuring bias is transferred from researchers to the participants, which is more consistent with bottom-up approaches to multiactor problems. level MCAs mainly address the latter issue via spatially aggregative methods, such as weighted linear combination, that are fully compensatory (i.e. increased performance at one location compensates for decreased performance at another location). However, the overall performance of a landscape management alternative cannot be reduced to the sum of performances of its spatial units at a given moment. Such additive aggregation to combine attribute maps ignores the existence of spatial relationships and patterns (Malczewski, 2000).
Two directions appear for reducing this gap. The first "corrects" the aggregation result by adding other evaluation techniques. This option is illustrated by Fürst et al. (2013), who use landscape metrics to improve assessment of esthetic and biological integrity services. The second acknowledges that "gaps" between the multiple levels of assessment are irreducible; as a consequence, levels of assessment should be explicitly distinguished. Though formal methods exist to identify gaps between organizational levels (e.g. MuSIASEM, Giampietro et al., 2009), relying on management levels or stakeholders' levels of interests can be more operational (see point iv). Using different criteria to assess different levels of the system could however undermine the consistency and intelligibility of the evaluation system. López-Ridaura et al. (2005) define disciplinary-and scale-independent criteria for sustainability assessments: productivity, stability, reliability, resilience and adaptability. This idea gave birth to the MESMIS framework (Astier et al., 2012), in which farm, community and regional sustainability are evaluated with the same set of criteria but with different indicators. Similarly, Manoli et al. (2005) use response properties (reliability, vulnerability, and resilience) to assess the temporal variability of management performances. These options offer a generic multi-level framework but are at odds with a bottom-up definition of criteria.

(vi) Uncertainties are under-considered
Finally, the question of uncertainties remains under-considered, or at best is reduced to a sensitivity analysis of the weighting systems and flexible support tools. We assert that sensitivity to scenarios of future change (as in Linhoss et al., 2013;Manoli et al., 2005), to resolution and spatial aggregation units (Malczewski, 2000), and to group delineation, among others, can enrich the deliberative process.
Moreover, since judgments can change during the deliberative process, solutions to capture uncertainties may also lie in the participatory settings (Garmendia and Stagl, 2010)

CONCLUSIONS
We critically examined the place of sustainability assessments within multi-criteria methods, trends in methods that combine a participatory and spatial approach, and current practices for assessing landscape-management alternatives through a systematic review of the literature, combining quantitative textual statistics from large datasets and qualitative analysis of specific research.
Drawing upon the results of textual statistics, we suggest that conceptual assumptions constitute the primary discriminating element of multi-criteria methods and that methodological elements should be both adapted to those conceptual bases (i.e. the nature of the problem tackled), as well as operational questions (i.e. the decision context). We offer insight into multi-criteria methods that differ from the classic multi-attribute/multi-objective typology or aggregation procedures. We consider the dichotomies of substantive/procedural rationality and grounded/positivist scientific legitimacy to explain the diversity of methods and approaches. In this respect, though both are grounded in real-life problems, spatial sustainability assessments and participatory sustainability assessments differ in what they consider a "rational" decision-process. Spatial assessments focus on accumulating as much information as possible to provide manager(s) with "all the elements" to make a substantively rational decision. In contrast, participatory approaches recognize deliberation as the principal justification of results. Consequently, cross-approaches are not so common.
When attempting to classify participatory spatial MCA methods, we noted that to assess a real situation, support design of a management alternative or accompany selection of an alternative, a wide variety of research postures towards society can be adopted. Unlike treatment of the participatory aspect, treatment of the spatial aspect depends to some extent on the MCA goal. For landscapemanagement issues, the method chosen should capture the diverse ways people relate to spaces and locations. Scenario-selection methods have the greatest potential to accommodate spatially-explicit, spatially-implicit, place-specific, overall visual and non-spatial values. However, putting this into formalized to better address the spatiality of problems. As a result, current cases of landscape-level MCAs resemble "patchworks" of different methods with distinct underlying hypotheses.
The main unaddressed issue of landscape-level MCAs is strikingly the question of distribution and heterogeneity. How are benefits distributed in space, time, and between social actors? How do spatial patterns and temporal behavior influence the overall "performance" of a management alternative?
These questions remain under-considered. It is implicitly assumed that increased overall value is desirable, irrespective of social or spatial justice issues. Similarly, most studies assume that the performance and acceptability of a management alternative only depends on the total extent of each land cover or land-use type, irrespective of its shape, location, or configuration. Assessments are grounded in a weak conception of sustainability and a reductionist view of the landscape. While MCAs provide formal methods to overcome problems of vertical aggregation (how to aggregate results for different criteria), horizontal aggregation (how to aggregate results from different locations) is rarely considered. In addition, although (or maybe because) more complex aggregation algorithms exist, they are not accessible or transparent to the wider public. Other important aspects of MCA for landscape-management issues remain under-investigated: how stakeholders are involved, the group values produced, the maps and other spatial representations valued, and the uncertainties addressed.
To conclude, we emphasize that MCA at the landscape-level do not succeed in addressing the joint issues of spatiality and multiple views, and thus has important scientific loopholes to close. MCA is nonetheless a promising method to structure landscape problems, to explore management options and to foster social learning. Even with these objectives, it is necessary to reveal biases and ethical positions implicitly assumed in the assessment process.