Seltz et al Evaluating light-based geolocation for demersal fisfies in high latitudes 



573 



surface, its location was determined by the Dop- 

 pler shift in the transmitted radio frequency in 

 successive uplinks (Keating, 1995). The end- 

 point position was the first location class (LC) 

 estimate reported in the LCl-3 range, which all 

 have error estimates <1.0 km. 



The basis of light-based geolocation is the 

 estimation of times of sunrise and sunset. 

 Two proprietary programs developed by the 

 Wildlife Computers, Argos Message Processor 

 (AMP, vers. 1.01.0007) and Time Series Pro- 

 cessor (TSP, vers. 1.01.0008), were used to ex- 

 tract times of sunrise and sunset from light 

 intensity data. AMP identified daily sunrise 

 and sunset times from light data transmit- 

 ted through Argos satellites or directly from 

 complete archival light records. TSP could 

 be used only to identify sunrise and sunset 

 times from complete archival light data from 

 PAT tags that were physically recovered. 



In the next phase, another Wildlife Com- 

 puters program. Global Position Estimator 

 (GPE, vers. 1.01.0005), used the sunrise and 

 sunset times to calculate the daily longitude 

 and latitude of tags. First, we rejected days 

 with light level curves that did not exhibit 

 smoothly sloping light levels from high to 

 low or low to high (Fig. 1). GPE was used to 

 calculate longitude for the remaining data 

 based on the local noon time of the tag (mean 

 of the sunrise and sunset times). Estimated 

 longitude values that were not possible for 

 a fish released in the Gulf of Alaska were 

 rejected from the data set. For example, an impos- 

 sible longitude was one that placed the tag on land 

 or outside the published range of the Pacific halibut 

 (i.e., to the west of Hokkaido. Japan (140°E) or to the 

 east of Santa Barbara, CA (117°W; Mecklenberg et al., 

 2002)). Once longitude was estimated, latitude was 

 estimated by GPE, which used the "dav/n and dusk 

 symmetry method" (Hill and Braun, 2001; Musyl et 

 al., 2001). Daily latitude estimates were the theoreti- 

 cal location of expected light levels that best matched 

 the observed light levels measured by the tag. Lati- 

 tude outliers were removed in the same manner as 

 that used for longitude outliers. For all three experi- 

 ments, the number of days with geolocation estimates 

 was defined as the days that produced latitude and 

 longitude estimates, after "bad" light curves (Fig. 1) 

 and outliers were removed. 



For the tags with known positions in the tank and 

 mooring experiments, we calculated bias and error 

 magnitude based on true locations. Daily positional 

 bias was calculated as the true position minus the 

 estimated position (signed distance between posi- 

 tions), and daily error magnitude was the absolute 

 value of the bias (distance between points). For the 

 tank experiment, we pooled the data from the two 

 tags. Mean error magnitudes of software types were 

 compared by using a two-tailed Ntest. For the fixed 



140 



120  



100 



80 



^,^»^- 



f-Ali^ 



'~'-^^vH, 



40 



20 





 0:00 



' 140 - 



120 - 



100 



80 



60 - 



40 



20 



2 00 4:00 6 00 8 00 10 00 12:00 14:00 16:00 18:00 20:00 22:00 00 



B 



A^-jA-J'-J^-M'- 



W 



iJ^^t 



.'VKjs,. 



P\ 



■^^^,J V.wvvjJv_ 



0:00 2:00 4 00 6 00 8:00 10:00 12 00 14:00 16:00 18:00 20:00 22:00 00 



Time (AST) 



Figure 1 



Examples of "good" and "bad" light curves. "Good" light curves 

 I A) have smoothly sloping sunrise and sunset events. "Bad" light 

 curves (B) do not have smoothly sloping sunrise and sunset 

 events and produce outlying longitude and latitude estimates. 

 The "good" light curve is from 2 March 2001 and the "bad" light 

 curve is from the same tag on 10 March 2001. AST = Alaska 

 standard time. 



mooring experiment, we calculated mean positional 

 bias and mean error magnitudes for each tag and 

 software combination. Mean biases were compared 

 to a hypothetical bias of zero by using a two-way 

 (tag and software) ANOVA model (vers. 8, proc GLM, 

 SAS, Gary, NC). Mean error magnitudes were com- 

 pared by using an ANOVA with a Tukey-Kramer 

 test (Kramer, 1956; vers. 8 proc GLM). For both bias 

 and error magnitude, the means are a measure of 

 accuracy and the standard deviations are a measure 

 of precision. 



For wild fish, it was impossible to know the true daily 

 position of each fish for the duration of the experiment. 

 However, for three of the eight tags released on wild 

 fish, geolocation estimates were produced in the first 

 or last six days of deployment. Therefore, we compared 

 the estimated positions of the tags for the six days im- 

 mediately following release of the tags and for the six 

 days before recapture of the tags or before tags trans- 

 mitted data to Argos satellites. All three of these tags 

 were physically recovered and TSP produced estimates 

 for all tags. AMP produced plausible estimates for one 

 tag only because other estimates were rejected as outli- 

 ers. For each comparison, we calculated the mean bias 

 and mean error magnitude, assuming that the fish was 

 stationary (or nearly so) during the first and last six 

 days of the deployment. 



