Domeier et al.: Tracking Thunnus thynnus onentalis with the aid of an automated algorithm 



303 



~J August 



September 

 7J October 

 ~J November 



December 

 ~J January 



Figure 9 



Positions for eleven Pacific bluefin tuna {Thunnus thynnus orien- 

 talis) tagged with satellite pop-up tags from 2000, 2001, and 2002 

 showing the 100% minimum convex polygon for fish positions within 

 a given month. 



been possible to be certain that this rapid excursion was 

 authentic without the aid of SST matching (as was also 

 done by Itoh et al. [2003a]). 



The PSAT Tracker algorithm provided relatively quick 

 and automated geolocation estimates for data recovered 

 from three separate types of tags deployed on Pacific 

 bluefin tuna. Furthermore, the PSAT Tracker latitude 

 solutions compared favorably to the light-based latitude 

 estimates during non-equinox times of the year. The use 

 of SSTs to resolve latitude allowed for spatial analyses 

 of individual bluefin positions for every month of the 

 year, whereas a strictly light-based approach would not 

 provide reliable latitude position estimates for approxi- 

 mately 30% of a year-long track. PSAT Tracker also 

 results in a global, rather than serial track solution. 

 In essence this means that no single position estimate 

 is selected without regard to the influence this position 

 has on the overall track. A serial track is one that is 

 produced by selecting each position without regard to 

 the effect each selection has on the overall track. A se- 

 rial track is also heavily biased by the start point and 

 may weight the location estimates based upon the pre- 

 vious location estimate, allowing a single poor location 

 estimate to ruin the remainder of the location estimates 

 for the track. 



It is instructive to compare our SST matching algo- 

 rithm to the Kalman filter-based algorithm developed 

 by Sibert et al. (2003). The Sibert et al. algorithm 

 depends solely upon light data collected by the tag to 



estimate latitude and longitude, whereas the PSAT 

 Tracker algorithm depends upon the light field to pro- 

 vide an estimate of longitude and solely upon the sea 

 surface temperature to provide an estimate of lati- 

 tude. The initial estimates of both approaches are then 

 refined according to a goodness-of-fit criterion that 

 depends upon assumptions regarding the swimming 

 behavior of the tagged fish. In the case of the Sibert 

 et al.'s algorithm, the behavior of the fish is modeled 

 in terms of a biased random walk model that describes 

 the movement of the fish in terms of an advection- 

 diffusion equation; the advective term describes the 

 most probable displacement of the fish during a time 

 step and the diffusive term describes the distribution 

 of less likely displacements. The usefulness of the ran- 

 dom walk model is largely determined by the adequacy 

 of describing the distribution of swimming speed and 

 direction of the fish. The algorithm also includes for- 

 mulations of the dependence of the accuracy and pre- 

 cision of the estimates of latitude and longitude from 

 the tag upon other factors. For example, around the 

 equinox the weighting of the estimate of latitude from 

 the tag measurements is greatly reduced (specifically 

 an inverse cosine squared function of date.) The Sibert 

 et al. algorithm simply searches for a track that mini- 

 mizes discrepancies between the positions predicted 

 from random walk model (the transition equation) and 

 those predicted from the tag measurements (the mea- 

 surement equation). 



