Accounting for uncertainty in marine ecosystem service predictions for spatial prioritisation

Spatial assessments of Ecosystem Services (ES) are increasingly used in environmental management, but rarely provide information on the prediction accuracy. Uncertainty estimates are essential to provide confidence in the quality and credibility of ES assessments for informed decision making. In marine environments, the need for uncertainty assessments for ES is unparalleled as they are data scarce, poorly (spatially) defined, with complex interconnectivity of seascapes. This study illustrates the uncertainty associated with a principle‐based method for ES modelling by accounting for model variability, data coverage and uncertainty in thresholds and parameters.


| INTRODUC TI ON
The Ecosystem Services (ES) concept reflects the goods and services provided by nature on which our societies rely for, among others, the provision of food, material, fresh water, climate regulation, soil formation, cultural heritage, and recreation (MEA, 2005) and provides a means of connecting complex ecological processes and functions to the benefits and values people derive from the environment.
The multidisciplinary nature of ES allows for a common language to express these benefits and values that people can extract, utilise or experience.In the past two decades, the ES concept has gained traction around the world with scientists and practitioners alike and has become influential in decision-making (Nahuelhual et al., 2015).
ES have found many applications, including spatial planning, ecosystem accounting, restoration initiatives, protected area design and hazard and risk assessments (Ruckelshaus et al., 2015).Many of these require a spatially explicit understanding of ES (Nahuelhual et al., 2015), and there are now many tools available (e.g.InVEST, ARIES, SolVES) that generate maps of ES as part of their assessment.
However, maps have an 'air of authority' in decision-making (Hauck et al., 2013), and are often assumed to be 'true' without validation or consideration of uncertainties existing within the models or tools from which they were derived.
Uncertainty estimates are essential to allow for confidence in the quality and credibility of ES assessments, but many challenges exist that need to be overcome to allow for the wider uptake of uncertainty analysis for ES (Hamel & Bryant, 2017).In terrestrial systems, a growing number of studies have focused on the impact of uncertainty in ES assessments originating from data quality (e.g.Snell et al., 2018;Westfall et al., 2021), model type or structure (e.g.Schmidt et al., 2016;Willcock et al., 2020), model parameters (e.g.Chisholm & Wintle, 2012;Grêt-Regamey et al., 2013), and spatial scales (e.g.Wang et al., 2018).However, in marine environments uncertainty in ES assessments is rarely studied (but see Cabral et al., 2017;Shi et al., 2009), and ES assessments in general are lagging far behind their terrestrial equivalents.Although better tools and models for marine environments are emerging, seascapes remain data scarce, poorly (spatially) defined, with complex spatial and temporal interconnectivity (Townsend et al., 2018).By extension, these factors create arguably higher uncertainty in marine ES assessment, which is rarely considered in modelling studies and their implementation in marine management and decision-making approaches like Marine Spatial Planning (MSP) or Ecosystem-Based Management (EBM).
For marine ecosystems, the needed shift from managing single species or sectors (e.g.fisheries management) to more holistic, ecosystem-based approaches has been recognised (McLeod & Leslie, 2009).MSP is a practical tool to assist implementation of EBM and requires information on the services generated in the environment to enable consideration of different uses and conservation of ocean space (Galparsoro et al., 2021;Santos et al., 2019).When trade-offs between activities, (biodiversity) conservation, and other management concerns exist, spatial prioritisation approaches can provide insight into protection of areas of highest value with lowest cost.Two methods that could be considered for spatial prioritisation include Marxan (Ball et al., 2009) and Zonation (Moilanen et al., 2005).Both methods can deal with ES as input data and define areas of highest priority for the assets of interest, but are more commonly applied for protected area design (e.g.Galparsoro & Borja, 2021;Leathwick et al., 2008) or restoration (e.g.Adame et al., 2015;Barbosa et al., 2019) in marine systems.For example, Zonation develops a land/seascape wide priority ranking for conservation, starting with the assumption to protect everywhere and progressively dropping sites after iteratively re-ranking sites whose loss has the lowest impact on protecting biodiversity or ES, and results in a nested priority rank map (Lehtomäki & Moilanen, 2013).
Accounting for uncertainty in marine ES provision forms a major gap in spatial prioritisation applications at present, with no known examples illustrating how uncertainty in ES predictions alters spatial prioritisation outcomes.Instead, studies on biodiversity conservation have applied spatial prioritisation models with and without uncertainty in species distributions (e.g.Langford et al., 2009;Stephenson et al., 2021).Both the distribution of the species themselves and the associated uncertainty vary spatially, allowing for highest priority areas to be identified and selected areas important/ suitable for conservation regardless of uncertainty.Likewise, both ES and uncertainty estimates are expected to be heterogeneous in space, meaning that not all areas contribute equally to ES nor are they equally certain (Grêt-Regamey et al., 2013).This could shift the focus to prioritise areas with high ES potential with high certainty and de-emphasise areas with low certainty in management and spatial planning to avoid negative surprises, similar to conservation applications (Moilanen et al., 2006).
In this study, we examined how uncertainty in ES predictions may impact MSP in marine environments.More specifically, we studied model uncertainty in four ES, representing major ES in shallow coastal areas (Smaal et al., 2019), provided by estuarine bivalves: recreational Food provision; Water quality regulation; Nitrogen removal and Sediment stabilisation.Services provided by two New Zealand bivalves, Austrovenus stutchburyi and Paphies australis, were modelled through a principle-based approach as per (Rullens et al., 2022).Uncertainty was assessed in relation to the density of bivalves and parameter settings in ES models to assess confidence in model predictions and determine how uncertainty discounting could affect priority areas for management.We hypothesise that uncertainty will be ES-and species-specific, given the species-specific bivalves, Ecosystem-Based Management, estuary, mapping, sensitivity analysis, spatial planning patterns and spatial heterogeneity observed in the original ES models.Furthermore, we expected that by accounting for uncertainty, spatial prioritisation will better capture highest priority areas for multiple ES.

| Study area and species
The effect of uncertainty on ES predictions and spatial prioritisation was assessed for a case study in Tauranga Harbour, New Zealand.
The study area encompasses a large (approx.200 km 2 ), mesotidal estuary of which 66% was intertidal habitat.Two key bivalve species in the harbour that contribute to ES are the infaunal and suspension feeding bivalves Austrovenus stutchburyi (littleneck clam) and Paphies australis (pipi) (hereafter Austrovenus and Paphies).Both species form high density beds with Austrovenus occupying mid to lower intertidal sand and mudflats, whereas Paphies predominantly occupying fast-flowing channels (Powell, 1979).Austrovenus and Paphies are the remaining dominant bivalve species in Tauranga Harbour, as many other species have declined or collapsed.Recognising the contribution of these species to the functioning of the system, the ES they provide, and their cultural significance make them an integral part of environmental management.

| ES mapping approach
A principles based approach (Townsend et al., 2011) was adopted to determine the contributions of Austrovenus and Paphies to Food provision, Water quality regulation, Nitrogen removal, and Sediment stabilisation (Rullens et al., 2022).In brief, specific sets of principles (summarised in Table S1) were designed for each ES that described the relationship between bivalve-mediated processes, functions and ES, reflecting the mechanisms by which ES are generated (Rullens et al., 2022).For example, the Food provision ES was underpinned by three principles, reflecting that food provision from recreational gathering is highest in (i) sites with high densities of large individuals, (ii) easily accessible locations for recreational gathering, and (iii) uncontaminated sites (Figure 1).Bivalve density is a foundational principle for all four ES, given many bivalve-mediated processes are density-dependent (Newell, 2004).Therefore, high-resolution (100 m) spatial predictions of the occurrence and density of Austrovenus and Paphies were generated from Species Distribution Models (SDMs) by modelling their relation with 6 environmental variables (biophysical and sediment characteristics) using Boosted Regression Trees(see Supplemenatry Information for detailed methodology; Rullens et al., 2021).Remaining principles reflect the environmental conditions best suited to support the provision of each ES within estuarine and coastal environments.
For each ES, principle scores were calculated using normalised scoring structures (ranging from −1 to 1) underpinned by thresholds identified in published literature (Figure 1).These scoring structures outline the change in principle score along increasing gradients of density or environmental variables.For each principle, a weighting was used to reflect the relative importance of a principle compared to other principles for that ES.Weightings ranged from 1 to 5 for any given principle and all principles per ES totalled 10.For example, for the ES Food provision, density of large individuals was weighted the highest (at 5), followed by accessibility with a weighting of 3 and heavy metal contamination with a weighting of 2. Scoring structure and weighting were applied to spatially-explicit density or environmental data layers in the harbour to create principle layers, which were summed per ES to create a total ES score out of 10.The original ES predictions and model settings are hereafter referred to as the baseline scenario.

| Sensitivity analysis
The uncertainty originating from SDM and ES model outputs and settings were assessed in a sensitivity analysis (SA) to determine their impact on ES predictions.Three types of uncertainty were considered, namely the uncertainty in bivalve density predictions generated by the SDMs, ES model parameter settings for weights and scoring thresholds (Figure 1).Two measures of spatially explicit uncertainty from the SDMs in Rullens et al. (2021) were considered for the sensitivity analysis.
Firstly, the range of density predictions for each species (per gridcell) was used as a representation of model prediction uncertainty.
That is, SDMs were developed using a bootstrapping method so as to generate 100 models trained on a randomly selected 75% of the available data and tested on the remaining 25%.These models were predicted to the study area to generate estimates of spatially explicit density at each bootstrap.In the ES baseline model, the mean predicted density was used by averaging over the 100 model predictions per gridcell.Here we used the range in predicted densities sampled in equal intervals (deciles) along the full range from lowest to highest predicted density for each gridcell, resulting in 10 density scenarios for the SA.Other dispersal metrics that capture the predicted range, such as standard variation, would likewise be suitable.
Gridcells with low uncertainty in the SDMs will have a small range in density predictions, whereas for high uncertainty gridcells density predictions would vary more widely.
Secondly, uncertainty can originate when SDMs are projected to new environmental conditions that were not covered in the training dataset.Environmental coverage is a metric that reflects how well the environmental conditions in an unsampled gridcell were represented by the sampled sites (Smith et al., 2013;Stephenson et al., 2020).Environmental coverage was modelled using Boosted Regression Trees (BRTs) for sampled sites (n = 156 for Austrovenus and n = 170 for Paphies) and randomly selected unsampled sites (n = 1000), in relation to the 6 environmental variables used in the SDMs (Rullens et al., 2021).BRTs with a Bernoulli error distribution were fitted to the 'present' sampled sites and the 'absent' unsampled box 2, Section 2.3) and Uncertainty  analysis (box 3, Sections 2.4 and 2.5).sites, with a learning rate yielding approximately 2000 trees and interaction depth of 2. This model was then applied to the study area identifying sites ranging from poor to perfect (0 to 1) understanding of the environmental space.Environmental coverage was then normalised from 0.5 to 1 and multiplied with the density scenario's from SDMs as described above.Thereby downweighing locations with low environmental coverage values as being less certain, compared to those with higher environmental coverage values being more certain (Stephenson et al., 2021).
For the ES models, uncertainty was considered for the threshold values in the principle scoring structure and the weighting values for the ES layers.All baseline threshold parameters were varied ±20% at 5% intervals around the original value, thereby creating 9 new threshold values (Figure 1, Table S2).For example, if the bivalve density threshold was originally set at 200 ind m −2 , the threshold in the SA would range from 160 (−20%) to 240 ind m −2 (+20%).
Depending on the shape of each principle's functional relationship, increasing or decreasing the threshold can have a positive or negative effect on the ES score.Nine scenarios for the SA were created that captured either all positive or all negative effects along the new threshold gradient (Table S1).Finally, sensitivity in weightings was assessed by consecutively changing the baseline weightings by ±1 point per ES, keeping the sum of all principle weights per service to 10, and no principles weights of zero (Rullens et al., 2022;Townsend et al., 2014).For an ES with three and four principles, this resulted in 7 and 11 unique weights combinations, respectively (Table S3).
For the sensitivity analysis, all combinations of scenarios for density, thresholds and weights were propagated through the model to calculate new ES predictions, resulting in 630 (ES with three principles) and 990 (ES with four principles) final combinations.To understand where the uncertainty originated from, the change in ES from the baseline (as the response variable) for each of the 630/990 scenario's per gridcell was regressed (using a generalised linear model -GLM) against scenario identifiers for density, thresholds and weights.For density, the difference between mean density (from baseline models) and the scenario density value was used.Thresholds were assigned a value for the percentage change in value (−20 to +20) and weightings were assigned as factors.The GLM included all three explanatory variables and their respective interactions.Variance was partitioned using the sum of squares from the GLM and expressed as a percentage to the total sum of squares.This process was repeated for 1000 randomly selected gridcells and averaged to estimate overall contribution of density, thresholds, and weights to changes in ES predictions from the baseline.

| Uncertainty in ES predictions
Outcomes from the sensitivity analysis were compared to baseline predictions to assess the accuracy of the original models.For each gridcell, the 5th and 95th percentile from the predicted distribution of the sensitivity analysis were used to summarise the uncertainty in the range of new ES values as the difference between baseline and lower/upper bound predictions.The 5th percentile was then carried forward as a conservative estimate of ES value (ES C ) in spatial assessments as it provides a value most unlikely of being too high.
Conservative estimates can capture negative surprises (i.e.areas that may provide lower ES value than was originally predicted) when compared to baseline ES models, and pose a risk in management.
Conservative ES estimate were subtracted from the baseline ES predictions resulting in an uncertainty value per gridcell (Figure 2).
Here, gridcells with a narrow range of newly predicted ES values in the SA will result in a small difference and thereby reflect areas of low uncertainty (Figure 2a,c).On the other hand, gridcells with a wide range of ES values from the SA could create large differences between the baseline and conservative values that reflect areas with high uncertainty (Figure 2b,d).Uncertainty values were assessed in accordance with the baseline ES value in uncertainty bi-plots, distinguishing areas of low (0-5), medium (5-8), and high (8-10) baseline ES values in combination with low (0-1), medium (1-2), and high (>2) uncertainty.

| Accounting for uncertainty in spatial prioritisation
To account for uncertainty in multiple ES, a spatial prioritisation analysis was conducted using the software Zonation (version 4) (Moilanen et al., 2005).Here, Zonation was applied on all ES to identify representative areas of highest ES potential by prioritising gridcells of highest value while simultaneously accounting for representation and complementarity, and aimed to maximise representation with the smallest possible cost, here area.The Zonation approach initially assumes that the entire study area is protected and sequentially removes cells with lowest contribution using the Core Area Zonation (CAZ) algorithm.When a cell is removed using the CAZ algorithm, the proportion of each layer located in each remaining cell goes up.This means Zonation tries to retain core areas of all layers until the end of cell removal (Moilanen, 2007).For each species, two Zonation analyses were run; one for the baseline ES predictions and one for the conservative ES predictions from the SA as a way of accounting for spatially explicit uncertainty in the prioritisation.Given the lower values for the conservative scenario, high ES area will be smaller and areas with originally high ES values but with low certainty should be de-emphasised.For both scenarios, all four ES layers were equally weighted and default parameter settings were used (e.g.edge removal, no aggregation algorithm, no cost layers, no administrative unit analysis, etc.).Zonation outputs included maps for spatial prioritisation, here focussed on the top 5%, 10%, 20% and 30% (nested) priority area for multiple ES potential.Furthermore, the Zonation outputs were used to study the proportion of the medium-high (>5) scoring areas for each ES contained within each prioritisation level to illustrate how well each ES is represented.Although conservative areas will represent a subset of areas of the baseline because of the de-emphasising of areas with low certainty, the uncertainty discounting analysis aims to avoid prioritising erroneous areas (most likely those areas with higher uncertainty) and may therefore provide a better outcome for multiple ES when accounting for uncertainty.

| Uncertainty in ES predictions
Uncertainty from the sensitivity analysis predictions showed wide ranges of uncertainty across ES and species, with maximum differences up to 5.8 out of 10 points on the lower bound (5th percentile) and 6.33 points on the upper bound (95th percentile) for Paphies Food provision (Figures S1 and S2).Although the largest differences were found for Food provision ES for both species, the distribution of the uncertainty were heavily right skewed (e.g.Figures S3 and S4 for lower bound), with the average uncertainty among the lowest at approximately 1-1.5 points difference with baseline ES values (Table 1).Instead, Water quality regulation had highest mean uncertainty on the lower bound for both species but lowest on the upper bound.This in particular indicated that baseline ES values for Water quality regulation might be an overestimate, given that the baseline prediction in some cases was in the top 95th percentile (shown as small negative values).While Nitrogen removal had among the lowest average and maximum uncertainty, it had a relatively normal distribution (Figures S3 and S4).Variance partitioning showed that overall  Note: Main finding for uncertainty partitioning were included for the effects of density, threshold and weights, see Table S3 for full results.
the weighting of principles had the highest contribution to explaining the uncertainty in ES values, ranging between ~25% and >50% of variance explained (Table 1, Table S3).Density had high contribution for some ES, notably Paphies Food provision and Nitrogen removal.
Biplots illustrated important spatial patterns in uncertainty and ES values which varied for the different ES and species of interest (Figure 3).Food provision ES showed large areas of high ES value with high certainty (in green), most notably for Austrovenus in the northern part of the harbour (Figure 3a).For Paphies, the high-density areas at the harbour entrances tend to support medium to high ES scores at high certainty (Figure 3b).However, for both species, the edges of the originally predicted high ES areas were found to be most uncertain (orange).Water quality regulation had few areas of high ES with high certainty, although some pockets remained for Austrovenus (Figure 3c).For Paphies, no high ES values were predicted in the baseline scenario (Table 1) and almost all medium scoring area were found to be highly uncertain (light orange).Nitrogen removal had large areas of marginal scores (medium score with either low or medium uncertainty) for both Austrovenus and Paphies (Figure 3e,f).
There were only few small patches of high uncertainty.Likewise, for Austrovenus Sediment stabilisation, most areas were marginal (low or medium baseline scores with medium uncertainty).Few patches across the harbour entrances had high scores, but all with high uncertainty (Figure 3g).

| Uncertainty in spatial prioritisation of ES
The top 30% priority area for Austrovenus baseline scenario were broadly spread throughout the northern and southern part of Tauranga harbour, with little/no contribution in the central part (Figure 4a).Some core areas were visible at sites across the northern and southern harbour entrance, notably for the top 5% and 10% areas (in dark red and pink respectively; Figure 4a).Further up the estuary, priority areas were more patchily distributed.For Paphies baseline scenario, the high-density beds at both harbour entrances form most of the core area in the baseline scenario, with few small patches elsewhere (Figure 4b).When accounting for uncertainty, many commonalities were identified in the locations for priority areas for both species (Figure 4c,d), although there was a shift in the level of priority with more top 5% and 10% area at the northern harbour entrances for both species.For Paphies, the main patches at the harbour entrances have shrunk and more small patches of top 20% and 30% (orange and purple) were visible further up the estuary (Figure 4d).
The percentage of medium and high ES value (>5) area that fell within the priority areas (top 5%, 10%, 20% and 30%) was calculated for both Baseline and Conservative scenarios (Table 2).Using conservative values from the sensitivity analysis reduced area of medium and high ES potential (Table 1), which affected how well these areas could be represented in the priority areas.By de-emphasising low certainty areas, the Zonation optimisation can capture the most certain and high value area for all four ES simultaneously (aligned to dark green zones; Figure 3), thereby representing a larger percentage of the available area for multiple ES in the conservative scenario.For example, 55.6% of medium-high Food provision area by Austrovenus was captured in the top 5% in the conservative scenario, compared to 29.1% in the baseline scenario.For some ES, the conservative scenario covered 100% in the top 20 and 30% priority areas.Food provision is captured best for both species and was generally between 2 and 3 times higher than other ES, but could be up to 6 times higher.
Nitrogen removal and Sediment stabilisation had lowest relative cover.
For Paphies Water quality regulation, no medium or high ES area remained in the conservative scenario.Austrovenus Nitrogen removal was the only ES that showed a small decrease in cover (0.2%-0.3%) at the top 20% and top 30% priority level.

| DISCUSS ION
The growing demand for spatially explicit estimates of ES provision in spatial planning is putting increasing pressure on models and tools to generate ES maps.However, overreliance on maps without information on the accuracy or certainty of predictions can introduce errors, affecting the quality and credibility of ES assessments and thus their utility for management.In this study we illustrate the uncertainty associated with a principle-based method for deriving ES estimates by accounting for model variability, model data coverage, and uncertainty in thresholds and parameters.There are various other factors from which uncertainty could arise, such as new environmental conditions, different modelling algorithms or variations in specific parameters, which could likewise be considered (Elith & Graham, 2009).While we selected the lower 5% bound to calculate uncertainty, the sensitivity analysis provides a breadth of ways to use the predicted ES distribution and choose metrics that best reflect project question or goals.Notably, using the top 95% of the predicted ES range would allow for benefit analyses or identify opportunities through 'positive surprises' (i.e.areas that may provide greater ES value than predicted by the baseline ES models) (Moilanen et al., 2006).The analyses presented here showed the overall good with results here suggesting that the single best way to improve ES models would be by improving our understanding of weights through expert elicitation.

| Insights for management and spatial planning
Uncertainty estimates provide important information for best management practices of ES.While ES are only one component of information needed in environmental decision-making, they allow for a range of benefits and values to be included.ES information can be used to inform management and spatial planning, but one must be cognisant of the cost of getting it 'wrong'.Costs include not only the investment of limited time, effort, and resources, but should also consider lost opportunity cost and loss of confidence in the approach or investments for restoration or conservation projects when they do not deliver as expected (Carwardine et al., 2008;Duarte et al., 2020;Schröter et al., 2014).Information from uncertainty analysis, as presented here, can provide decision-makers with insights to help reduce the risk of prioritising the wrong area or function as a first step before further work is undertaken to identify potentially valuable areas for which there is (currently) uncertainty.
Uncertainty assessments can promote transparency allowing decision makers to define acceptable confidence levels, or providing information that would change the recommended course of action (Bryant et al., 2018)

F
Illustration of uncertainty calculations for hypothetical distributions of sensitivity analysis (SA) values per gridcell (ES k,j ).The distribution (black) shows the frequency of newly predicted ES values in the sensitivity analysis, with the baseline ES value plotted in green and the conservative value from the SA distribution (as the lower 5th percentile) plotted in red.The Uncertainty (Unc i,j ) measure is illustrated as the difference between the baseline and conservative value in panel (b).Four scenarios (a-d) are depicted for low-high baseline ES value in combination with low-high uncertainty.TA B L E 1 Summary of lower (5th percentile) and upper (95th percentile) bound uncertainty (mean and maximum) for each ES per species.
Finally, Paphies Sediment stabilisation had originally large areas of medium and high scoring potential, which was reduced in the biplots to core zones in the high-density beds at the harbour entrances (Figure 2h).Sediment stabilisation maps showed substantial areas of marginal (intermediate) combinations, including examples with high ES values with medium certainty (brown) or medium ES values with high certainty (light green).
performance of the principle-based approach to ES modelling, given low average uncertainty values for individual ES and overall robust spatial prioritisation outcomes when accounting for uncertainty in multiple ES assessments.However, considering model uncertainty brought to light important nuances in priority areas with implications for ES management.As expected, variability in the magnitude and spatial distribution of uncertainty was observed between individual ES and differed between the two species.The magnitude of uncertainty differed slightly but consistently, with higher maximum uncertainties for Paphies than Austrovenus for all ES.Between species, ES pairs showed comparable patterns in mean and maximum uncertainty.For both species, Water quality regulation had highest average uncertainty, whereas Food provision had highest maximum uncertainty.The spatial distribution of uncertainty showed that the core highdensity patches had among the highest certainty for medium and high scoring ES.Fringes of high-density patches were found to be more uncertain as they can be transition zones for one or more environmental variables.Finally, ES with low representation of high ES patches (e.g.Nitrogen removal and Austrovenus Sediment stabilisation), were highly uncertain (dark orange), while all other areas for these ES were marginal or low.Uncertainty estimates can thereby contribute to guide further investigation, including empirical validation of high-uncertainty areas of interest or those with low representation.Furthermore, the variance partitioning can improve understanding of which aspect caused uncertainty and requires further attention, F I G U R E 3 Ecosystem service and uncertainty biplots for Austrovenus stutchburyi (left) and Paphies australis (right).Ecosystem services were modelled for Food provision (a, b), Water quality regulation (c, d), Nitrogen removal (e, f), Sediment stabilisation (g, h).Biplots display ES scores from the baseline model as low (ES score < 5), medium (ES score 5-8), and high (ES score > 8) in combination with uncertainty from the sensitivity analysis as low (0-1), medium (1-2), and high (>2).F I G U R E 4 Spatial ES prioritisation for Austrovenus stutchburyi (left) and Paphies australis (right).Spatial prioritisation analyses were conducted for baseline (a, b) and conservative (c, d) scenarios.Areas were categorised from the highest to lowest priority (top 5%, 10%, 20% and 30% areas).
The spatial prioritisation analysis using Zonation took a different approach to this problem, running analyses on the baseline and conservative estimates, accounting for all four ES simultaneously.By doing so, areas of high ES and low certainty will have reduced values compared to the baseline predictions, but by comparison may still exceed areas of medium ES potential (with high certainty).The spatial prioritisation in this study showed similar core areas identified in the harbour, but with shifting priority levels.Firstly, this implies that the principle-based approach for ES modelling is reasonably robust against uncertainty.Decisions based on priority areas from the baseline scenarios would likely still result in important areas being captured in the top 30%.However, the nuance comes from the shift in classification of top 5 and 10% area, towards the high-density beds in the northern part of the harbour.ES uncertainty biplots showed that these are predominantly areas of medium to high ES with high certainty for Food provision, which is further supported by the strong representation of Food provision area for both species.On the other hand, Nitrogen removal was covered less well in the conservative Zonation analysis, as it had large marginal areas and low representation of high ES(Rullens et al., 2022), that did not overlap well with other ES.By accounting for uncertainty, the area of medium and high scores was reduced and allowed for the spatial prioritisation to cover the areas of highest value more comprehensively.The spatial prioritisation in combination with the ES uncertainty biplots thereby provide important information for spatial planning for individual and multiple ES to focus on area of highest value with highest certainty.This type of information is urgently needed in marine ES assessments and for their management, which extends to other environments to improve transparency and confidence in the quality and credibility of ES assessments.TA B L E 2 Percentage of medium and high ES (>5) valued area captured within the priority areas (top 5%, 10%, 20% and 30%) for the baseline (B) and conservative (C) scenarios.