TRACKING WILDLIFE BY SATELLITE 



25 



10 



15 



20 



25 30 



35 



Sensor Counts 



Fig. 21. Short-term activity indices for a free-ranging bull elk 

 engaged in six activities. Data courtesy of D. Vales, University 

 of Idaho. 



message per minute and a maximum overpass length of 

 about 17 min.) For each activity type, the frequency of 

 each count from the calibration trials (see previous para- 

 graphs on work with captive animals) was divided by the 

 total number of counts observed during that activity, 

 yielding the count's "known" probability, given the ani- 

 mal's true activity. Sensor counts from each overpass 

 (from free-ranging animals) were similarly divided by the 

 number of counts obtained during the overpass. Thus, data 

 from each overpass also yielded a probability distribution, 

 but from an unknown activity type. To classify an over- 

 pass, we totaled the absolute values of the differences 

 between the two distributions for each activity type. A 

 perfect fit between the two distributions produced the 

 minimum score of 0; an observed distribution that lacked 

 any overlap with a known distribution yielded the max- 

 imum score of 2. Each overpass was then assigned the 

 activity type yielding the lowest score. Figure 22 provides 

 an example from two representative overpasses from a 

 caribou of the Central Arctic herd, showing overpasses 

 that were categorized as inactive and running behavior, 

 respectively. The implicit assumptions underlying the 

 classification algorithm were that successive counts from 

 the same overpass were independent, and that animals did 

 not change activities during an overpass. 



We hypothesized that errors in activity classification 

 were more likely to occur in shorter overpasses. To ad- 



dress the relation between length of overpass (number of 

 counts) and classification accuracy, we conducted 

 computer-based trials in which we randomly selected 

 counts from each distribution until the specified number 

 had been accumulated. We then subjected each series to 

 the comparison procedure described (in the preceding 

 paragraph), tallying the number of correct and incorrect 

 classifications. The results displayed the expected pattern: 

 higher classification accuracy was associated with in- 

 creasing number of counts (longer overpasses; Fig. 23). 



We estimated the expected magnitude of classification 

 errors and biases (Table 14) on the basis of mean overpass 

 length obtained during caribou studies in northern Alaska. 

 Assuming the activities of wild caribou were similar to 

 those of captive animals, the magnitude of error in behav- 

 ior classification was low. Feeding activity would be ex- 

 pected to be classified erroneously as inactive behavior in 

 approximately 5% of occurrences. Other misclassification 

 probabilities were generally <1%. 



In late 1987, the manufacturer examined the mercury 

 switches on PTT's recovered from caribou and found that 

 4 of 12 had hairline cracks that caused unreliable activity 

 data. An additional PTT had a mercury switch inclined at 

 an incorrect angle. For this reason, the analysis of ac- 

 tivities was restricted to only the seven PTT's known to be 

 operating correctly, effectively limiting the period of 

 study to October 1986-October 1987. 



Results. For all caribou combined, patterns of activity 

 budgets followed pronounced seasonal trends (Fig. 24). 

 Lying (or inactive), the most common behavior category, 

 peaked in April and declined each successive month until 

 reaching a low in July and August. Feeding was greatest in 

 November and December, with a secondary peak in July- 

 September. Walking and running occurred infrequently 

 but were far more common during the period of insect 

 harassment (July and August) than during any other 

 period. 



These patterns differed in some respects from Alaskan 

 caribou seasonal activity budgets estimated by previous 

 observational studies. In particular, the proportion of time 

 spent feeding, as determined from the PTT, was considera- 

 bly lower and the time spent inactive higher than doc- 

 umented by Roby (1978) and Boertje (1981). However, 

 the seasonal trends closely followed patterns in seasonal 

 movements described for the same herds by Fancy et al. 

 (1989). Among individual caribou, variability in apparent 

 activity budgets was substantial, particularly for June 

 (Fig. 25). We lack data to speculate how much this vari- 

 ability reflects true differences among individuals and 

 how much might be due to differences in the time of day in 

 which sampling occurred, fit of the collar on the animal, 



