Causal approach to environmental risks of seabed mining

Seabed mining is approaching the commercial mining phase across the world’s oceans. This rapid industrialization of seabed resource use is introducing new pressures to marine environments. The environmental impacts of such pressures should be carefully evaluated prior to permitting new activities, yet observational data is mostly missing. Here, we examine the environmental risks of seabed mining using a causal, probabilistic network approach. Drawing on a series of interviews with a multidisciplinary group of experts, 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 modelling to provide estimates of mortality of benthic fauna in the Baltic Sea. The model is used to evaluate 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, highlighting the importance of a cross-disciplinary 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.

for resource extraction (Hein et al. 2013). However, dealing with impacts of activities that 50 do not take place yet means that there is no observational data on the impacts, with high 51 uncertainties on both the implementation of the activity and its consequences for the 52 environment. This uncertainty creates a challenge to estimate the impacts in a way that is 53 scientifically robust, while accounting for the knowledge gaps and scarcity of data to 54 support decision-making. 55 Current plans for mining are outlined both in shallow continental shelf areas and the deep 56 sea, encompassing areas within national jurisdiction of sovereign states and the 57 international seabed in the 'Area' (Miller et al. 2018). While most initiatives are still at an 58 exploratory stage, the increasing need for raw materials is pushing countries to consider 59 where to get their mineral resources in the future ).  Environmental risk assessment (ERA) is a process aiming to evaluate the different 70 possible outcomes following human activities (Burgman 2005). A risk in this context is 71 defined as any unwanted event (here 'impact') and its probability. Currently, most ERAs

92
Bayesian networks (BNs) are graphical models that represent a joint probability distribution 93 over a set of variables and provide an alternative to commonly used scoring procedures in 94 ERAs (Pearl 1986;Kaikkonen et al. 2021). In BNs, the strength of each connection 95 between variables is described through conditional probabilities. As probabilistic models, 96 the result of a BN is not a single point estimate, but a probability distribution over the  Here, we describe an approach for integrating expert knowledge into a causal risk 105 assessment for seabed mining. We use the Baltic Sea as an example to test our 106 approach, as mining iron-manganese nodules has already been tested in an industrial 107 setting in this area (Zhamoida et al. 2017) and the ecosystem components and food web

109
Given the number of ongoing seabed mining initiatives and attempts to quantify impacts, 110 the aim of this work is to provide a framework that allows combining information from 111 multiple sources by bringing ecological information to risk analysis while explicitly 112 addressing uncertainty. To move towards a quantitative risk assessment, we demonstrate 113 the use of BNs in an operational setting and discuss needs for a quantitative ERA in the 114 context of emerging maritime activities.

116
Our case study deals with ferromanganese (FeMn) concretion removal in the northern 117 Baltic Sea. The Baltic Sea is characterized by low species richness compared to many 118 marine areas, and the food web structure and ecological traits characterizing major taxa 119 have been well described (Törnroos and Bonsdorff 2012). Due to the relatively shallow 120 depth of the Baltic Sea, the extraction activity is to some extent comparable to sand and

139
We apply a 3-step approach for working together with experts to create a model that 140 summarizes the causal connections in the system and enables providing quantitative risk 141 and uncertainty estimates (Fig. 2). The first step consists of mapping the relationships 142 between key drivers and ecosystem responses with experts in semi-structured interviews.

143
The use of structured methods for expert elicitation has been highlighted in recent years,    Semi-structured interviews were held at a location chosen by the interviewee or via an 173 online connection. For face-to-face interviews, causal maps were drawn on paper, 174 whereas in online interviews maps were constructed using an online drawing tool. All 175 interviews were recorded with consent from the interviewee.

176
At the beginning of each interview, participants were introduced to the use of causal 177 networks. Each expert was presented with the same scenario of the mining activity and the 178 changes in the environment arising from the activity, noted as pressures (Table 1) (Table 1). These 188 variables form the core of the model by describing the basic processes related to mining.

189
To explore the ecological impacts arising from these pressures, the following eight 190 interviews were conducted with marine ecologists. Each expert was presented with the 191 same scenario of the mining activity and the physicochemical pressures identified in the 192 first phase with the geologists (Table 1). The experts were then asked which ecosystem 193 components they think will be affected by these pressures. Whenever possible, experts  variable. For instance, the terms "polychaetes", "annelids", and "worms" were grouped 210 under 'mobile infauna' (see Table S1 in Supporting Information for full details of individual  where the term IM x(1-DM) accounts of the proportion of fauna remaining after direct 259 extraction. We applied numerical approximation at 1% accuracy to calculate joint 260 probabilities of the combined discrete classes (Table 2) for total mortality used in the 261 model.

262
The resulting CPTs were incorporated in the BN model created in R software (R 2020).  Table 2).

274
The modelling was done using R 3.6.3, with package bnlearn (Scutari 2009      The full causal model is highly complex (Fig. 3), and parameter estimation would be a  In the case of mining 75% of a discrete mining block, the most probable outcome in terms 348 of total mortality for both sessile epifauna and infauna is estimated to be 81-100% 349 mortality (Fig. 5, A). The probability of the highest mortality for sessile epifauna is slightly 350 higher than for infauna (60.1% compared to 57.7%, respectively). For mobile epifauna, 351 60-80% mortality is the most likely outcome with a 52.2% probability.

352
The likeliest outcome of the mining scenario described above in terms of indirect mortality

358
The BN model allows estimating the probability of any variable of interest in the model 359 (here relative mortality) given certain evidence (e.g. regarding the mining operation or 360 environmental conditions). To give an example, when mining occurs on only 50% of a 361 discrete block, but release of harmful substances is known to occur, the probabilities for 362 the indirect mortality of benthic fauna are higher for all groups (Fig. 5, B). These changes 363 illustrate the relative importance of certain pressures on the overall mortality.  Table S5 428 for spatial and temporal extent of the pressures), resulting in immediate impacts, chronic 429 and long-term impacts, and factors affecting the recovery potential of organisms. To support decision-making on potential future use of seabed resources and further Supporting information (SI S1-S5) are available as an attachment to this manuscript, as 508 well as at https://github.com/lkaikkonen/Causal_SBM .