Validation approaches of a geolocation framework to reconstruct movements of demersal fish equipped with data storage tags in a stratified environment
Introduction
Tracking of aquatic animals has provided unique insights into their behavior and distribution and has fundamentally altered our understanding of the structure and function of global aquatic ecosystems (e.g. Hussey et al., 2015). Data storage tags (DSTs) are among the wide range of technologies available to track movements of marine species. DSTs deliver time-stamped measurements of environmental parameters experienced by the tagged individual (e.g. ambient temperature, salinity, pressure, light level), and, more recently, individual parameters (e.g. acceleration, heart rate); however, these data often lack spatial context. Geolocation tools are regularly used to infer movements of tagged individuals by comparing spatially resolved oceanographic data from, for example, 3-D ocean models with in-situ environmental data recorded by the tags (Arnold and Dewar, 2001; Skomal et al., 2009; Thorstad et al., 2014; Le Bris et al., 2018) while accounting for characteristics of the individual fish movement, such as distribution and range of swimming speeds.
Historically, many geolocation tools have been developed for particular areas, species and tag types, resulting in several case-specific solutions. Most recent geolocation model frameworks have been based on non-parametric state-space models, such as the Hidden Markov Model (HMM) (Nielsen et al., 2019). Several applications of HMMs have demonstrated these approaches perform well for near-shore demersal species tagged with depth sensors (Andersen et al., 2007; Nielsen et al., 2019), including the baseline framework for many current HMMs that was originally developed for the geolocation of North Sea cod (Gadus morhua) using tidal information (Pedersen et al., 2008; Thygesen, 2009). This has since been applied to geolocate Atlantic cod (e.g. Thorsteinsson et al. 2012; Le Bris et al. 2013a; Liu et al. 2017) and has been adapted to a variety of other systems, mainly by adjusting the likelihood estimation to better account for the characteristics of different tags, species and areas (Le Bris et al., 2013b; Woillez et al., 2016; Biais et al., 2017; Braun et al., 2018a). Many of the original applications of the Pedersen et al. 2008 model framework were case-specific and written for use in Matlab, likely restricting wider adoption of these methods. However, Braun et al. (2018) generalised the core functionality to facilitate applications beyond specific species and regions and translated the base functionality into the open-source R language (R Core Team, 2017) as the R-package HMMoce (https://github.com/camrinbraun/HMMoce), thus widening the range of potential users and making it readily adaptable to other applications.
To apply a geolocation model to a new area, species and or tag type, it is essential to estimate the performance, classify the associated uncertainty and test for the robustness of the desired model framework and specifications (Nielsen et al., 2019). Otherwise, movements and behavior derived through sub-optimal geolocation techniques might be error-prone (Braun et al., 2015) and could lead to misinterpretation. Validation of geolocation methods is usually conducted by comparing modelled and known tracks. While data on known movements of tagged individuals is typically rare, some surface-oriented, epipelagic species have been effectively double-tagged with archival tags (e.g. DSTs, pop-up satellite archival transmitting tags) and satellite-linked location tags (e.g. salmon sharks (Teo et al., 2004); blue sharks (Braun et al., 2019a)). These double-tagging efforts can provide independent, known tracks for validation of geolocation techniques. Similarly, some studies have successfully double-tagged demersal species with a DST and an acoustic tag to supplement geolocation likelihoods with additional known positions from an acoustic receiver array (Liu et al., 2017). However, both double-tagging with satellite and acoustic tags are only applicable to a limited number of cases and are relatively costly (Arnold and Dewar, 2001; Block et al., 2005).
Other studies have used simulated tracks, generating potential natural behavior of the species of interest, for model validation. These simulated tracks can then be matched with oceanographic parameters which are used to subsequently reconstruct the tracks with a geolocation model (Neuenfeldt et al., 2007; Righton and Mills, 2008; Liu et al., 2017; Nielsen et al., 2019). This simple but effective method can emulate the present knowledge of fish behavior and helps to develop a better understanding of processes within the geolocation model (Jonsen et al., 2013) and the choice of input parameters. However, geolocation models are rarely tested in their ability to reconstruct stationary behavior, for example, by applying these methods to stationary, moored DSTs or temperature values recorded at an automated measuring station (Hunter et al., 2003; Thorsteinsson et al., 2012; Liu et al., 2017). Furthermore, testing the model’s ability to detect mobile behavior by attaching sensors to fishing gear to reconstruct the vessel’s track has, to our knowledge, only been presented in Righton and Mills (2008). These approaches have usually only covered small areas compared to the stock distribution area and simulated rather local behavior. Preferably, a combination of different validation methods across a spectrum of space and timescales should be conducted to account for different types of movement behavior (ideally for the target species), but these validation methods are not routinely used before applying a geolocation model to a new study species or area.
Previous validation studies found that a higher horizontal or vertical spatial heterogeneity in environmental parameters, and thus larger variations in water depth and temperature measurements of the DST, can lead to improved accuracy of the geolocation (Liu et al., 2017; Nielsen et al., 2019). DSTs implanted in cod in the southern Baltic Sea between 2016 and 2019 indicated high heterogeneity in the recorded temperature and depth data as tagged individuals made regular daily vertical movements across several meters of a stratified water column (Hüssy et al., 2020). Thus, the combination of dynamic oceanography and animal behavior can be used to further constrain the geolocation problem and should thus be leveraged to improve reconstructed tracks (e.g. Braun et al., 2018b).
Here, we adapt the existing HMMoce model framework to better handle geolocation of DSTs deployed on demersal species and quantify the accuracy of the adapted model to assess its utility for future studies geolocating tagged cod in the southern Baltic Sea. In particular, we tested the sensitivity of the model results to the vertical resolution of the regional ocean model used to calculate likelihoods, as well as swimming speed parameters, using five experiments: (1) artificial tracks, (2) stationary DSTs moored inshore, (3) an automated offshore measuring station, (4) temperature-depth probes attached to an otter board of a commercial fishing vessel, and (5) DSTs attached to the CTD probe and the otter board of a fisheries research vessel. We compared known positions to modelled outcomes from the adapted HMMoce model to assess the uncertainty of the geolocation model across a wide range of possible fish behavior and movement types.
Section snippets
Oceanography and cod behavior in the southern Baltic Sea
The bathymetry of the temperate, non-tidal southern Baltic Sea is characterized by shallow nearshore areas and deeper offshore basins which provide a high degree of heterogeneity in bathymetry (Fig. 1A). In each basin there are weak horizontal gradients in water temperature (i.e. within the same depth layer). A three-layer structure characterizes the water column in the Arkona and Bornholm Basin. This double stratification is composed of a permanent near-bottom halocline and a seasonal surface
Geolocation of simulated tracks
The general trend of the simulated tracks could be depicted correctly with the adapted HMMoce (Fig. 2d–f). Release and recapture position were modelled correctly when the maximum swimming speed was chosen equal to or above the constructed maximum speeds, otherwise the gaussian-shaped movement kernel prioritized smaller movements that the extent of movement required. Even choosing the maximum swimming speed multiple times higher than the modelled track led to the correctly constructed release
Discussion
To put data from archival tags such as DTSs into spatio-temporal context, geolocation tools must be adapted to and tested in new environments and for new applications. We have described five validation approaches which imitated the stationary and mobile behavior of cod in the southern Baltic Sea tagged with temperature-depth DSTs. Both simulated behaviors could be reconstructed with the adapted R-based geolocation model HMMoce. The uncertainty associated with the estimation of daily positions
CRediT authorship contribution statement
Stefanie Haase: Writing - original draft, Conceptualization, Formal analysis, Methodology, Software. Uwe Krumme: Writing - review & editing, Conceptualization. Ulf Gräwe: Writing - review & editing, Resources. Camrin D. Braun: Writing - review & editing, Software. Axel Temming: Writing - review & editing.
Declaration of Competing Interest
The authors report no declarations of interest.
Acknowledgements
We thank Ina Hennings and Anne Georgi (OF) for deploying and heaving the stationary DST buoy. We are also grateful to Kay Briesewitz, captain of SAS111 “Christina Bettina”, and Andreas Hermann (OF) for deploying the temperature-depth probe at the otter board and the BSH for providing the data from the MARNET-station. We also thank the captain of the FRV “Solea” and the cruise leader Andrés Velasco for attaching DSTs at the CTD and the otter board and Lena von Nordheim for revising an early
References (43)
- et al.
State-space models for bio-loggers: a methodological road map. Deep Sea Res. Part II top
Stud. Oceanogr
(2013) - et al.
Supplementing electronic tagging with conventional tagging to redesign fishery closed areas
Fish. Res.
(2013) - et al.
Transequatorial migrations by basking sharks in the Western Atlantic Ocean
Curr. Biol.
(2009) - et al.
A HMM-based model to geolocate pelagic fish from high-resolution individual temperature and depth histories: european sea bass as a case study
Ecol. Model.
(2016) - et al.
Using the particle filter to geolocate Atlantic cod (Gadus morhua) in the Baltic Sea, with special emphasis on determining uncertainty
Can. J. Fish. Aquat. Sci.
(2007) - et al.
Electronic tags in Marine fisheries research: a 30-year perspective
- et al.
Vertical movements of a pelagic thresher shark (Alopias pelagicus): insights into the species’ physiological limitations and trophic ecology in the Red Sea. Endanger
Species Res.
(2020) - et al.
Return migration patterns of porbeagle shark (Lamna nasus) in the Northeast Atlantic: implications for stock range and structure
ICES J. Mar. Sci.
(2017) - et al.
Electronic tagging and population structure of Atlantic bluefin tuna
Nature
(2005) - et al.
Movements of the reef manta ray (Manta alfredi) in the Red Sea using satellite and acoustic telemetry
Mar. Biol.
(2015)