Abstract
One way to try to quantify displacement mortality, and sources of variation in mortality, is via a mechanistic model. Simulations from the mechanistic model can be used to estimate the mortality associated with different scenarios, and these results can then be used (WP4) in estimating displacement mortality rates.
We focus here upon using SeabORD, an individual-based model of seabird behaviour, energetics, demography and windfarm interactions, to estimate the displacement mortality rates associated with different colonies under different wind farm scenarios. SeabORD includes two main mechanisms of windfarm interaction: displacement effects (a switch in foraging location as a result of a windfarm) and barrier effects (increased flight distances, and hence energetic costs, to reach foraging locations when birds avoid flying over windfarms). We refer to the combination of both of these as “displacement mortality”. SeabORD can consider either a single windfarm, or the simultaneous effects of multiple windfarms, and considers impacts on both chicks and adults. SeabORD directly simulates the impact of windfarms on chick mortality. For adults, SeabORD simulates the impact on windfarms on the change in body mass over the course of the breeding season and then translates this into impacts on annual adult survival via published mass-survival relationships. SeabORD only considers the impacts of windfarm interactions during the chick rearing period within the breeding season and is parameterized for four species: black-legged kittiwake, common guillemot, razorbill and Atlantic puffin.
The biological parameters within SeabORD are largely fixed based on expert judgement or published values from the literature (Searle et al., 2018). However, the remaining inputs to SeabORD - bird distributions maps, colony sizes, and prey maps – need to be specified separately for each colony. In addition, there is one input parameter for SeabORD – total prey – that cannot meaningfully be derived from expert judgement or the published literature. SeabORD outputs are highly sensitive to the values of this parameter, so it is calibrated separately for each colony by selecting the parameter values for “total prey” that lead baseline adult and chick survival rates (in the absence of a windfarm) to lie within biologically plausible ranges. Because the values of this parameter are calibrated against a range of baseline survival rates, a range of plausible parameter values are identified for each colony. Parameters from across this range are then used when simulating windfarm impacts via SeabORD, allowing this key source of uncertainty to be accounted for within the SeabORD outputs.
SeabORD simulates changes in adult and chick mortality (and survival) rates, and in a range of other quantities: adult mass loss during the chick rearing period is one of the key outputs, because the adult mortality and survival rates are derived from this using published relationships. In principle, we could use SeabORD to provide model-based estimates of displacement mortality for a wide range of colonies and windfarm scenarios, and (via the work in WP4) thereby build up information on typical mortality rates, and on the variability in these rates between colonies and windfarms. In practice, SeabORD is a computationally intensive model to run, so there would be substantial practical challenges in doing this. The recoding of SeabORD into R as part of the Marine Scotland CEF project, which we exploit here, has involved improvements to computational efficiency, but realistically complicated runs of SeabORD continue to require large amounts of computer time. In addition, the calibration process within SeabORD has always required human input, and this is required each time the model is run for a new colony for which it has not previously been calibrated. Work within the CEF project to try to automate the calibration process revealed additional challenges in doing this – the work aiming to find simple proxies of SeabORD outputs that could be used to predict whether SeabORD would produce broadly plausible baseline demographic rates for a particular total prey level, but the results suggest that the obvious choices for such proxies did not perform well in predicting whether SeabORD would produce plausible baseline demography. For the moment, SeabORD therefore continues to require a manual calibration step for each colony.
Emulation provides a framework for using statistical models to approximate mechanistic models. It uses a “training set” of mechanistic model inputs and outputs to build a general model for the relationship between the mechanistic model inputs and outputs, and, as such, provides an approximation to the mechanistic model. Emulation is typically designed to approximate computationally intensive models using much less computationally intensive models, so that the emulator can then be used, predictively, as a rapid but approximate substitute for the mechanistic model. As SeabORD is a computationally intensive model it is an obvious candidate to be emulated. The types of models used for emulation are generally similar to the models that could be used to model relationships in empirical data – for example, multiple regression, mixed models, Gaussian processes, random forests and neutral networks. Within the context of emulation, the “response variables” are the mechanistic model outputs (or a subset of these), and the “explanatory variables” are the mechanistic model inputs (or a subset of these).
Emulation is not a replacement for mechanistic modelling: the emulator is designed to provide a more rapid alternative to the mechanistic model, but at the cost of some loss of accuracy caused by using the emulator to approximate the mechanistic model. This accuracy can be increased by using a large training set (i.e., more mechanistic model runs), but since the rationale for using the emulator is to reduce computational effort, there is a trade-off: more mechanistic model runs will lead the emulator to produce a more accurate approximation to the mechanistic model but will take more computational time. In the extreme, using an extremely large set of mechanistic model runs would allow us to build an emulator, but would also negate the main rationale for using the emulator (computational savings). Using a very small number of mechanistic model runs as a training set would represent a major computational saving, but would risk producing an emulator that was a poor approximation to the mechanistic model.
Aside from computational savings, another rationale for using the emulator is that it allows us to examine the properties of the mechanistic model. In particular, it allows to examine the extent to which variations in the outputs from the mechanistic model can be explained by simple relationships between the model inputs and outputs. This is useful in determining the key features of the mechanistic model that are influencing the behaviour of the model and may be useful in identifying parts of the model that could be simplified without loss of accuracy.
Within this work package we focus upon using SeabORD, and an emulator of SeabORD, to produce estimates of displacement mortality for three species – kittiwake, guillemot and razorbill – under to a range of SPAs and windfarm scenarios, and to identify sources and levels of variation in these mortality rates within each species. For each species we focus upon attempting to build an emulator of SeabORD wind farm impacts that can be applied to all UK SPAs, and, for each SPA, to all wind farms whose footprints are in the CEF Data Store and that lie within the foraging range of the SPAs. This represents an exceptionally large number of scenarios, so we build the emulator using a much smaller training set of SeabORD model runs for each species. For pragmatic purposes, we focus upon three SPAs per species and use colony-specific bird distribution maps derived from the maps of Wakefield et al. (2017) and upscaled to SPA level within the CEF. Since the vast majority of possible windfarm scenarios involve extremely low levels of interaction between windfarm footprint and SPAs, and practical interest lies in situations in which there is a fairly substantial interaction with windfarms, we develop the training set using windfarm scenarios whose interaction with an SPA (as defined using “totalpinords”) exceeds a minimum threshold, and thereby consider the effects of between 5 and 11 windfarm scenarios per SPA per species.
We use the SeabORD training runs to build an emulator for each species, in which we relate the SeabORD impacts on adult and chick mortality, and adult mass change, to a range of metrics that summarise the characteristics of the windfarm footprint(s), SPA and spatial interaction between SPA and windfarm(s). We use this emulator to identify the percentage of variation in impacts that can be explained by these characteristics, and to identify the key characteristics that influence the simulated impacts for each species. The aim of the emulator is to predict the levels of displacement mortality that we would expect SeabORD to produce for a much wider set of SPAs and windfarm scenarios than those used in developing the emulator.
The results of the work must be interpreted cautiously, in large part because of the relatively small training set of SeabORD runs that we were able to use to build the emulator for each species, so we conclude by outlining the key limitations and caveats underpinning the work, and describing the potential for future work in this area (including work that is already planned to take place within WP4 of the ECOWINGS project). We finally conclude by examining the wider implications of the work.