Combining Predicted Seabird Movements and Oil Drift Using Lagrangian Agent-Based Model Solutions

Abstract

In traditional oil spill risk assessments, the mortality of seabirds is typically assessed based on a simulated amount of oil combined with a statistical and static (seasonal mean) number of birds within a given grid cell. The size of the cell is typically in the order of 10 by 10 km. Cell averaging in a coarse Eulerian grid will inevitably introduce a high degree of uncertainty with respect to real impact, and due to the patchiness in seabird distribution may result in over-estimation of impacts outside high-density areas and underestimation within high-density patches. Lagrangian agent-based modelling of species movements and oil drift directly would provide consistent results independent of the grid resolution and, at the same time, provide a fine-scale resolution of potential impacts. The robustness of this approach is demonstrated for a potential oil spill in the Barents Sea in an area with a high density of Common Guillemot, followed by a discussion on how this approach can improve future risk assessments during oil spills.

1. Introduction

In traditional risk assessments, the mortality of seabirds is typically assessed based on a simulated amount of oil combined with a statistical and static (seasonal mean) number of birds within a given grid cell; the size of the cell is typically in the order of 10 by 10 km [1]. It is obvious that cell averaging in a coarse Eulerian grid introduces a high degree of uncertainty with respect to real impact, and due to the patchiness in seabird distribution may result in over-estimation of impacts outside high-density areas and underestimation within high-density patches.

As an alternative to the Eulerian approach risk assessments may be undertaken using the concept of Lagrangian particle tracking. The Lagrangian approach potentially improves the accuracy of risk assessments as the movement of oil particles can be simulated in parallel with a simulation of density and movement of seabirds using the same weather and oceanographical model scenarios. In addition, the predicted impact will have high spatial precision as both oil and seabird particles will be resolved independently from any model grid. With the Langrangian model approach the industry standard risk assessments of oil spills [2, 3] could be further improved. Additionally, the approach would make the results of oil risk assessments in different geographical areas more comparable in the future.

The applicability of this approach in future oil risk assessments is demonstrated for a potential oil spill in the Barents Sea based on the results of the Marine Animal Ranging Assessment Model Barents Sea (MARAMBS) project.

The MARAMBS project (2018–2019) was funded by the Research Council of Norway and aimed for improving the knowledge of the distributional dynamics of seabirds in the Barents Sea. The tentative oil spill scenario was undertaken for an area with a high density of Common Guillemot (Uria aalge) during 30 days of the post-breeding period in September 2016.

2. Lagrangian modelling concept

Both the oil spill and the bird densities are simulated using a Lagrangian particle modelling approach. Each particle represents generally a set of variables and attached computations. This means that instead of simulating the distribution/concentrations per grid element (Eulerian model space), the density/distribution emerges as the result of particles moving across the model domain (Lagrangian space). Note that particles can move freely within the model space, i.e. they are described by a continuous point location instead of stepwise grid coordinates. Parameters like flow speeds, water depth etc. are interpolated from the computational mesh elements to the point coordinates of a particle. The differentiation between a Lagrangian oil spill and an agent based model (ABM) for the seabirds is more conceptual as both use the very same underlying techniques; an ABM model usually describes particles/agents that control their motion based on some decision-based rules computations whereas the movement in simple Larangian models is mostly passive, i.e. current driven.

2.1 Agent based modelling (ABM)

As described above the movement of an oil spill particle model mainly governed by the underlying hydrodynamics, wind forcing and weathering process. The sea-bird movement is the outcome of a complex agent-based model (ABM) describing the movement, feeding and internal states based on a set of input factors (forcing) and sub-models as outlined in Figure 1 and Table 1. The model template used is the CBIRD ABM module in DHI’s MIKE ABM Lab framework [4].

Figure 1.

Conceptual diagram of the various parameters/sub-modules affecting movement mode decisions in the envisioned seabird ABM. Green box variables denote Eulerian spatiotemporal model forcings, while blue boxes indicate Lagrangian variables/processes. Red arrows indicate two-way feedback mechanisms.

Overview of CBIRD modules with brief descriptions
Module	Description
Core movements	Once sea-birds leave their colonies after the breeding season, they will switch between 3 states: resting on the water surface, swimming on or below the water surface, and flying in the air.
Bioenergetic Module	Governs physiological processes such as thermoregulation, energy expenditure of different core movements and digestion.
Day/night	Sea-birds can be set to observe diurnal behavioural cycles.
Habitat memory	Location of past encounters with preferable habitat can be readily accessed by the sea-bird. This increases overall foraging success.
Ice/land	Ice and land both represent obstacles that a sea-bird must avoid if it is to remain in comfortable environmental conditions. This module aids navigation around land masses and encroaching ice.
Wind	Unobstructed on the open ocean, sea-bird movements are directly affected by wind drift when sea-birds are on the water’s surface or flying. Wind speeds and direction also affect take-off probabilities.
Social	Sea-birds exhibit flocking behaviour, and tend to move in groups to feed, avoid predation, or react to local environmental conditions.
Storm	Stormy weather can threaten sea-bird survival, especially with no shelter on the open water. Sea-birds will often move away from this perceived threat.
Moult	Some species cannot fly while moulting at sea. Besides limiting their excursion range, the start of moulting is also critical to ensure that sea-birds are not initially stranded in poor habitat patches.
Moulting area	Once leaving their colonies for moulting areas, some species are observed to be extremely targeted in their trajectory towards these moulting areas. They will fly over good habitat patches and only begin moulting in these specific areas, mainly in the Central Barents region.

Table 1.

Overview of CBIRD modules.

One of the main challenges related to agent-based modelling of the behaviour of individual sea-birds is to strike a balance between realistic parameterisation and heuristic representativeness of the model.

The CBIRD ABM model is described in detail in [5]. The description is repeated in brief below. It largely follows the ODD protocol for describing individual- and agent-based models [6, 7]. The ODD protocol consists of seven elements: the first three elements (purpose, entities, overview) provide an overview, the fourth element (design concepts) explains general concepts underlying the model’s design, and the remaining three elements (calibration, parameterisation, validation) provide details.

The purpose of the CBIRD model is to predict the dynamic spatiotemporal distribution of seabirds like the Common Guillemot in the Barents Sea by combining several individual movement and feeding behaviours as a function of explicitly modelled bioenergetics with the included effect of physical environmental forcings, such as ocean currents, wind drag, habitat suitability and ice cover.

The model tracks the horizontal position and internal state of individual seabird agents (the Lagrangian entity) inside the spatially explicit model domain. Each agent represents multiple individuals and can thus be considered a ‘super-individual’ [8]. The memory of previously visited habitat locations (x, y coordinates) is stored as attributes of the entity, while the strength of the habitat memory is described by a reference memory (dimensionless) to the previously visited habitat versus a satiation memory (dimensionless) relating to the perceived habitat utility (dimensionless) of staying in the current habitat. The time attribute, time since the last food encounter (minutes), controls the magnitude of swimming activity of migrating birds relative to drag forces imposed on agents by wind and currents.

All model calculations of state variables and updates of environmental forcings occur at a discrete time step over the simulation period. At the beginning of each model time step, the following sequential order is applied:

1. Update of environmental input forcings. For environmental forcings of different spatiotemporal resolution than the ABM model, spatiotemporal interpolation routines are applied

2. Calculation of process equations in the Eulerian Framework

3. Update of state-variables in the Eulerian Framework

4. Update of sensing functions for each agent relative to both Eulerian and Lagrangian frameworks

5. Calculation of Lagrangian arithmetic expressions based on updated values obtained from (1), (3), and (4)

6. Update of Lagrangian state-variables based on calculations listed in (5)

2.2 Oil spill modelling

The oil model is a multi-component model, i.e. the total oil mass is distributed among different hydrocarbon components, typically defined by their density. The masses and the properties of each particle may change over time due to weathering. Each oil component has its own weathering process. The fate of the spilled oil is typically divided into different processes:

• Drifting - (the motion of the oil caused by the ambient winds, currents and waves)

• Spreading (Figure 2) the motion of oil induced by its buoyancy and surface tension properties relative to water. Oil spilled on the water surface immediately spreads over a slick of few millimetres in thickness. The spreading is primarily promoted by gravity and surface tension; however, many spills of varying sizes quickly reach a similar average thickness of about 0.1 mm.

• Weathering (Figure 2) processes causing physical and bio-chemical changes of the oil by evaporation, emulsification, biodegradation, dispersion, dissolution, oxidation, and sedimentation and beaching. Spreading, evaporation, dispersion, and dissolution can be defined as short-term weathering processes, whereas emulsification, biodegradation, and photochemical oxidation are recognised as long-term weathering processes (Figure 2).

Figure 2.

Processes acting on spilled oil, (modified after [9]).

2.3 Interaction between agents

A risk screening is carried out for Common Guillemot based on the oil spill model results by computing the spatial overlap of the oil slick and the simulated species distribution. Each oil particle represents a given area with a given slick thickness and relevant properties such as oil viscosity and depth below the water surface. The latter determines whether the oil floats on the surface or the oil droplets are dispersed into the water column due to wave action.

For each time step in the oil spill simulation period, it is tested if the position of the individual sea bird agent particle is within the oil area represented by the individual oil spill particles. If this is true, the sea bird agent particle is flagged as being in contact with oil. However, it is further tested before flagging if the following parameters exceed threshold levels for the sea-bird in question.

• Diving depth (Zmin). If the oil particle is below Zmin, there is no flagging.

• Oil slick thickness. If the oil slick thickness is lower than the threshold level, there is no flagging.

• Viscosity (cP). If the viscosity is higher than the threshold value, there is no flagging.

Each sea bird particle/agent represents a large number of individuals, i.e. it is a so-called super individual. This must be considered when assessing whether or not individuals of a species have been in contact with a drifting oil slick. In case of super individuals, the point position of the agent/particle is replaced with an area around the particle. It is assumed that the individual birds represented by the super individual particle are uniformly distributed within this area. If there is a spatial overlap with the area defined by any oil particle, the affected number of individuals corresponds to the relative overlap area of the bird species particle. The area covered by the bird species particle is then reduced proportionally or, when no unaffected individuals remain, the particle is removed from the analysis. This means that the number of affected birds serves as a worst-case estimate (Figure 3).

Figure 3.

Schematic of the calculation of the affected number of individuals based on the overlap area of the oil- and bird species particle.

The risk screening for the oil spill is then expressed in terms of the number of birds exposed (flagged) to oil above the threshold values during the oil spill simulation period. The number of birds exposed is also expressed as a fraction of the average population size with the model domain during the oil spill simulation period. The oil spill simulation will typically continue 30 days after the spill has terminated.

The above approach is based on pure Lagrangian ABM results for birds and oil, respectively, which in turn provide consistent results independent of the Eulerian grid mapping (i.e., results that are independent of the grid resolution).

3. Case study in the Barents Sea

The hydrodynamic data and most forcing parameters for the ABM and Oil spill model are driven by DHI’s existing 3-dimensional hydrodynamic model. The model domain, encompassing the Barents Sea and the colonies in Norway, Russia and the islands of Svalbard, Franz Josef and Novaya Zemlya is shown in Figure 4. The modelling software package applied was DHI’s 3-dimensional flexible mesh model, MIKE 3 FM [10] which includes both meteorological, tidal and oceanographic effects.

Figure 4.

Common guillemot modelled mean density September 2013. Pink dots indicate common guillemot survey sightings. The red circles/ellipses encompass the main colonies. Model domain extent in stereographic projection (Stereo_1670_Sc1).

3.1 Common guillemot

Common Guillemot is a large auk species with a circumpolar distribution, breeding in dense colonies between 40°N and 75°N. Like other auk species, the wings are used for both swimming and diving. Common Guillemots are excellent divers with maximum diving depths of more than 200 m. It is a pelagic pursuit diver, and the primary food item in the Barents Sea is small pelagic fish such as capelin and fish larvae and fry. Common Guillemots leave the colony in late July - early August. The male follows the chick for the first two months at sea. During this period, they moult and are flightless for approximately 45–50 days.

The population of Common Guillemots in the Barents Sea has declined dramatically since the first censuses in the 1960s. In particular, the mass mortality of birds during the winter of 1986–1987 decimated the populations along the Norwegian coast and on Bjørnøya. This incident was probably caused by food limitation due to a crash in the capelin stock and very low densities of alternative food items. Since then, the populations have been growing, but the species is still listed as critically endangered on the Norwegian Red List. Colonies are found in the southern part of the Barents Sea. The Barents Sea populations are year-round residents to the southern Barents Sea area, and in addition, Common Guillemots from Norwegian colonies further south migrate into the Barents Sea during autumn.

The population of Common Guillemot of 300,000 pairs is concentrated in colonies around Bjornoya/Bear Island (83.9% of the population) and the Finnmark Coast (8.9%), with minor colonies at the Murman Coast (5.4%) and in the Norwegian Sea (1.8%). *https://www.npolar.no/en/species/common-guillemot/

3.2 Oil spill

For illustration, the oil from a hypothetical oil spill in the Barenst Sea was applied as a representative oil type with a spill rate corresponding to 700 m³/day and a spill duration of 8 days. The spill is assumed to be a topside surface spill.

The spill was released on 1st September 2016 at (73° 34′ 40″ N, 22° 55′ 5″ E), and the simulation continued 30 days after the spill had terminated.

3.3 Mortality assessment

The sensitivity of different sea-bird species to oil pollution is described as a function of the relative amounts of time spent on the sea surface, the importance of the population within the study area and the size of the world population. The knowledge underpinning these parameters is relatively well established due to large amounts of standardised data from countrywide surveys of beached dead sea-birds, surveys of sea-bird densities at sea, counts of the number of breeding sea-birds in colonies and information on specific details of breeding and survival.

Thresholds of relevant species of sea-birds in the Barents Sea to the characteristics of oil residues (key effect features) were provided. The key effect features of oil in relation to sea-birds are film thickness, viscosity, and depth of the oil in the water column (Table 2). Although the effect in relation to depth in the water column can be confidently assessed from general knowledge of the diving capacity of the different species, the actual diving depth per species may fluctuate depending on season and location. It may differ in the Barents Sea from other areas. More importantly, the thresholds related to film thickness and viscosity have been established without undertaking tests on live birds [11]. Hence, they should be regarded as guidelines rather than well-defined thresholds.

Species	Zmin (m)	Oil slick thickness	Viscosity (cP)
Glaucous Gull	<20	>10 μm	<3000
Black-legged Kittiwake	<20	>10 μm	<3000
Common Guillemot	<100	>10 μm	<3000
Brünnich’s Guillemot	<100	>10 μm	<3000
Atlantic Puffin	<50	>10 μm	<3000
Little Auk	<30	>10 μm	<3000

Table 2.

Example of effect features of oil residues relevant for assessing impacts of oil on sea-birds in the Barents Sea. Zmin is the depth at which the species will be at risk from oil.

The modelled distribution of Common Guillemots showed peak occurrence at the boundary between the Norwegian Coastal Current and the Atlantic water mass. The predicted (mean) distribution of suitable Common Guillemot habitat displays a well-defined concentration 150–200 km north of the Norwegian coast. Both oil and Common Guillemots displayed a high degree of spatio-temporal variation which resulted in a limited intersection between the modelled oil slick and the main concentration of the guillemots (Figure 5). The intersection took place in the area just south of Bjørnøya. An animation of the 38 day simulation period can be displayed here: Video: https://www.youtube.com/watch?v=lgzBrfiQm6g.

Figure 5.

Overlap between modelled mean oil density (L/km²) and modelled mean density (n/km²) of common guillemot during the test period. - “sources: Esri, DigitalGlobe, GeoEye, i-cubed, USDA FSA, USGS, AEX, Getmapping, Aerogrid, IGN, IGP, swisstopo, and the GIS user community”.

The estimated hourly number of impacted Common Guillemots is displayed in Figure 6. Due to the limited intersection between the oil and guillemot particles the guillemots were only impacted during three of the 40 days in the model period (2–4 September 2016). Further, the modelled impact only took place during 15 of the 213 modelled hours equivalent to 7% of the trajectory time. The total number of impacted guillemots during the modelling period was 50,000. With a total number of breeding pairs of 300,000 the number of Common Guillemots in the western Barents Sea during autumn 2016 would be approximately 1 million, taking juveniles and non-breeding immatures and adults into account. Thus, the relative impact on the population represents 5% of the population (Figure 7).

Figure 6.

Estimated hourly number of casualties of individual common guillemot during the oil spill simulation 1/9–8/102016.

Figure 7.

Cumulative impact of simulated oil spill on common guillemot population in the Barents Sea 1/9–8/102016.

4. Conclusion/discussion

This demonstration has stressed the potential for applying a combined fine-scale Langrangian modelling design for improved prediction of the movements and intersections between an oil spill and seabirds. The oil spill risk modelling method provides a powerful tool for an initial screening of the potential impact on the various seabird species during a given period, even if the approach does not comply with a full risk assessment in line with the traditional MIRA class approach in Norwegian Waters [1]. Furthermore, combining the ABM model results of species movement directly with the result of oil particle movements provides consistent results independent of the grid mapping (i.e. results that are independent of the grid resolution). This makes results from different areas directly comparable.

More importantly, studies of the marine distribution of birds unambiguously point at the fine-scale distribution of most species of seabirds. Seabirds predominantly show an affinity to physical oceanographic properties such as fronts, upwellings and eddies, which enhance the probability of predators encountering prey [12, 13, 14]. In the Barents Sea this tendency is reflected in the ubiquitous concentrations of seabirds in the region of the Polar Front [15].

To accurately describe the overlap between the distribution of seabirds and oil slicks over time, one needs to be able to take account of the high degree of clustering and habitat association seen in seabirds. This solution can only be achieved with a Langrangian modelling approach like the one tested here. By using a standard Heuristic approach based on mean seasonal distrubutions of seabirds predicted impacts will unlikely resolve the true intersection in the distribution of the oil slick and the seabirds.

In other words, if high-resolution Langrangian models are not applied as part of the risk assessment of oil incidents mean values rather than in situ values for oil and seabirds predicted intersections will rarely match reality. As a result, risk assessments may lead to a type II error—a result estimating an impact in an area of low seabird density —or a type I error—a result erroneously pointing at a smaller or medium impact in an area where seabirds are highly concentrated. Thus, despite the large number of risk assessments of oil spills undertaken in the past accurate assessment of the impacts of oil slicks on seabirds remains a challenge.

The combined fine-scale Langrangian modelling design seems to have a strong potential to pave the way for more realistic assessments of the concurrent distribution and movement of oil slicks and sensitive species of seabirds. Obviously, the calibration of the both the oil spill and seabird models requires that detailed empirical data are available regarding i) the oceanographic properties of the spill site, ii) the chemical composition of the oil and iii) the local density and distribution of the target species of seabird. Spatially refined assessments of the risk of seabird species sensitive to oil pollutions will enhance both the planning and environmental management of oil and gas exploration activities in the future.

Acknowledgments

Special thanks to Equinor, ConocoPhillips and TotalEnergies for funding and valuable discussions during the development of the MARAMBS data portal and special thanks to the Norwegian Institute for Nature Research (NINA) for participation in the development of the data portal in relation to seabirds.