Causal Approach to Determining the Environmental Risks of Seabed Mining

Mineral deposits containing commercially exploitable metals are of interest for seabed mineral extraction in both the deep sea and shallow sea areas. However, the development of seafloor mining is underpinned by high uncertainties on the implementation of the activities and their consequences for the environment. To avoid unbridled expansion of maritime activities, the environmental risks of new types of activities should be carefully evaluated prior to permitting them, yet observational data on the impacts is mostly missing. Here, we examine the environmental risks of seabed mining using a causal, probabilistic network approach. Drawing on a series of expert interviews, we outline the cause-effect pathways related to seabed mining activities to inform quantitative risk assessments. The approach consists of (1) iterative model building with experts to identify the causal connections between seabed mining activities and the affected ecosystem components and (2) quantitative probabilistic modeling. We demonstrate the approach in the Baltic Sea, where seabed mining been has tested and the ecosystem is well studied. The model is used to provide estimates of mortality of benthic fauna under alternative mining scenarios, offering a quantitative means to highlight the uncertainties around the impacts of mining. We further outline requirements for operationalizing quantitative risk assessments in data-poor cases, highlighting the importance of a predictive approach to risk identification. The model can be used to support permitting processes by providing a more comprehensive description of the potential environmental impacts of seabed resource use, allowing iterative updating of the model as new information becomes available.


S1 Methods description
A causal approach to determining the environmental risks of seabed mining Step 1: Causal mapping with experts to define model structure The first stage of the work focuses on defining which ecosystem components should be included in the risk model and mapping out their interdependencies based on a previous network drawn from literature.
Model building consisted of first interviews with experts in geology and mining technology to recognize the pressures caused by mineral extraction and affected non-biological ecosystem components (e.g. changes in seabed topography, sediment composition), and then with experts in ecology to complete the pressures and the affected ecosystem components.
Each expert was presented the same scenario of FeMn concretion extraction (including e.g. water depth at extraction site and extraction technique), and was then asked which ecosystem components would likely be affected by the pressures resulting from extraction activities. This first round of interviews resulted in 11 individual causal maps detailing the pressures caused by mineral extraction, the affected ecosystem components, as well as a number of additional variables We note that the sufficient number of experts is context-specific, and very different numbers of participating experts can be found from the literature (see e.g. the review by Krueger et al 1 ). Clemen and Winkler 2 highlighted the diminishing marginal returns of large numbers of experts and suggested using 3-5 experts, whereas Morgan 3 , for instance, emphasized that if the experts share similar views of the science underpinning their understanding of the system, 5-6 experts is enough, but a larger group is needed if the experts express diverse opinions. In our work, we included new experts in a stepwise process until no further information (e.g. new variables) was introduced to the model. Hence, we adopted the "theoretical saturation" concept which suggests continuing sampling until information is simply confirmed but not modified or elaborated 1 .
While we believe that 11 experts formed an adequate sample in our study, we also recognize that this may not always be the case, especially when dealing with poorly known systems, where information on the different ecosystem components and processes may be very scattered to topic experts. However, it should be noted that the difference between poorly known and well-known systems is not straightforward, as the experts can disagree on relatively fundamental issues even in systems that are studied relatively thoroughly, like the Baltic Sea (see e.g. Uusitalo et al. 4 ).
Step 2: Finalizing the model structure The causal maps resulting from the first interviews were combined into a consensus map, which all participating experts had the possibility to comment on.

Adjacency matrices
The first step in constructing the combined causal map consisted of listing all different variables in the individual causal maps. We the evaluated which concepts described the same variable or process (e.g. "oxygen concentration" and "hypoxia") and removed any duplicate variables (see also section 2.2. on functional groups).
Using these same harmonized variable names for all the causal maps, we then coded the connections in each individual causal map into a combined adjacency matrix, which illustrates the connections between the pressures and the affected ecosystem components (see example in Table  1). To evaluate the views between experts, we summed the number of times a connection was mentioned in the interviews.
We had also asked the experts to rank the strength of the causal connection from 1-3 (3 being strongest), but due to the high number of connections in many of the maps, we were not able to elicit all the link strengths. For this reason, and for wishing to illustrate a complete picture of the different impact pathways, we ended up including all unique variables into the combined causal map. However, first summarizing the causal connections into tables using either average link strengths of the number of times the connection was mentioned in the interviews would be useful in cases where there is disagreement between experts as to the direction of the causal connection, or if there is a need to limit the number of connections in the model for e.g. further quantification of the model. In our case, we used the information on the link strengths when selecting the variables for the Bayesian Network model.

Functional groups
The final functional groups included into the model were combined of the taxa and functional groups directly mentioned in the interviews, and the trait expressions that were mentioned to define the vulnerability to the pressures caused by mineral extraction (e.g. feeding habit). An example of all the possible trait combinations for benthic fauna is presented below (Figure 1).
The number of trait expressions/modalities used to determine the groupings varied between functional groups. Although the theoretical number of random trait groupings can be potentially large, only a small subset of combinations are likely present and further considered in the combined network model. If a trait was not seen to affect the response in some broader group of organisms, all the different combinations of trait expressions were not used. Here, traits are treated as discrete variables for simplification, although in nature species may express a variety of traits. These functional groups have been drafted based on the expected acute response. Any additional traits affecting recovery potential could be added in further model development.

Review of model structure
To ensure that the combined map represented the views of the experts involved in the model framing, all participating experts had the possibility to comment on the model structure in an open online document. Given the large size of the model, the causal network was presented both in the form of a graph ( Figure S7) and a table (Table S4-S6). The document and the comments on the model structure were visible to the other experts to encourage knowledge exchange and learning between participants. The experts were also given the option to send their comments directly to the modeller.
Step 3: Bayesian Network building The final causal model was modelled as a Bayesian network (BN). Bayesian statistics provide an alternative to commonly used simple scoring procedures in ecological risk assessment. For an introduction to BNs, see e.g. [5][6][7] . The BN was modelled as expert system, meaning that no empirical data was directly incorporated in the model. The modelling was done using R statistical software, with package "bnlearn".
Based on the results of the expert interviews, we quantified only a sub-model of the complete causal network focusing on three groups of benthic fauna: sessile filter feeding epifauna, mobile epifauna, and burrowing infauna. The BN model was developed from variables describing these three benthic faunal groups, the main pressures affecting them, and any intermediate variables between them in the combined causal network. To reduce complexity of the model in terms of spatial and temporal dimensions of the impacts in the first stage of the work, we restricted the model to account only for the acute impacts within a spatially discrete mining block.

Defining discrete variable states
Discrete variable states were defined to describe the variation in the magnitude of pressures arising from the mining activity. This included considering which combinations of pressures the extraction and sediment deposition are likely to result in. The discrete states may be described through quantitative metrics, like different concentrations of substances or depth in centimeters, or can be based on qualitative descriptions of discrete classes, like high, medium, low. The key aspect with regard to the biological responses was that these pressure levels should make sense from an ecological perspective. As the demonstration in our case study is not bound to a specific setting, we decided to frame the model variables quite generally with the states low-moderate-high. The discrete variable states were mostly defined based on expert judgement, informed by literature.

Probability elicitation
Within a BN, the magnitudes of impacts are illustrated through conditional dependencies. The probabilities of each value of the child node, conditioned on every possible combination of values of the parent nodes, were drawn from expert opinion. These describe the strength of the causal relationships between variables in the model (Figure 2). We used the graphical interface provided in the ACE application 8 to initialise the conditional probability tables (CPTs) with one expert in geology and one benthic ecologist. The application provides a starting point for defining the overall shape of a conditional probability distribution by allowing ranking the direction and magnitude of the parent nodes on the child node and populating the table through a scoring algorithm. For the probabilities concerning the impacts of direct pressures on benthic fauna, the prefilled tables evaluated and adjusted in another session with another benthic ecologist to reach a consensus on the magnitude of the impacts.
Direct and indirect mortality were modelled separately, so that the total mortality of benthic fauna comprises both the direct mortality from the extraction of sediment and mineral concretions and the indirect mortality stemming from the other pressures from the extraction activity. This allows for estimating the effects of the pressures for both the direct mining area (total mortality) and in neighbouring areas (indirect mortality).
Direct mortality was estimated as a direct proportion of the mined area, so that e.g. mining 50% of an area results in 50% of fauna being extracted. For the remaining (indirect) pressures, we first populated the CPTs with one benthic ecologist using the ACE application and then refined them using Google sheets. To evaluate the first estimates, we organized another session with another benthic ecologist s who had been involved in the model building and was available for further work to finalize the CPTs on indirect pressures ( Table 2). We elicited the frequencies using the following questions:  What is the lowest the value could be?  What is the highest the value could be?  What is the most likely value? Finally, the separate CPTs for both direct mortality and indirect mortality were combined so that the total mortality of benthic fauna within a discrete block and one moment in time comprises the direct mortality from extraction of sediment and mineral concretions, and the indirect mortality of the remaining fauna that are exposed to the pressures from the extraction activity (i.e. one organism cannot be both extracted and die of sediment deposition). The probability of total mortality of benthic fauna was thus calculated as: where p(Indirect Mortality) x (1-p(Direct Mortality)) accounts of the probability of the proportion of fauna remaining after direct extraction. While the resulting joint probabilities are continuous, here we calculate them at 1% accuracy (e.g. round them up) to provide the probabilities of a mortality of a given proportion of benthic fauna for the five discrete classes we use in our model.   S14 Figure S3.6 Interview 6 with a marine ecologist. The colored ovals depict the pressures which were presented at the beginning of the interviews and served as a starting point for the causal mapping exercise. S15 Figure S3.7 Interview 7 with a marine ecologist. The colored ovals depict the pressures which were presented at the beginning of the interviews and served as a starting point for the causal mapping exercise. Numbers from 1 to 3 indicate the strength of the causal connection, 3 being the highest. S16 Figure S3.8 Interview 8 with a marine ecologist. The colored ovals depict the pressures which were presented at the beginning of the interviews and served as a starting point for the causal mapping exercise. Figure S3.9 Interview 9 with a marine ecologist. The colored ovals depict the pressures which were presented at the beginning of the interviews and served as a starting point for the causal mapping exercise. S17 Figure S3.10 Interview 10 with a marine ecologist. The colored ovals depict the pressures which were presented at the beginning of the interviews and served as a starting point for the causal mapping exercise. S18 Figure S3.11 Interview 11 with a marine ecologist. The colored ovals depict the pressures which were presented at the beginning of the interviews and served as a starting point for the causal mapping exercise. The different arrow strengths indicate the strength of the causal connection (thickest being strongest).