Gilpatnck: Estimation of dolphin school size with aerial photography 



643 



were considered too variable and little confidence was 

 placed in the accuracy of the estimate. 



Analytical methods The CV was plotted against the 

 variables "school size" and "quality rating" to evaluate 

 how different school sizes and photograph image quali- 

 ties affected precision. The CV was also used to char- 

 acterize the precision of individual readers in repeated- 

 counts as described below. 



The percent deviation (PD) is a measure of how dis- 

 tant (+ or -) a count is from the mean. Because this 

 difference is expressed as a percentage of the mean, 

 the PD is a consistent index of deviation relative to 

 changes in the mean. For the PD, let .r„ be the ith 

 reader's determination of the size of school j. The av- 

 erage determination (or mean) over three readers of 

 the size of thejth school was given by 



1 3 



3 U 



(1) 



The PD of x is expressed as 



PD„ = -A 



100 



(2) 



PD values were plotted to evaluate whether indi- 

 vidual reader counts were normally distributed or read- 

 ers were biased (i.e., tended to count high or low rela- 

 tive to the mean). To test for temporal trends in 

 individual reader counts, PD values were regressed 

 against the variable "time." Time represented the 

 chronological sequence, unique for each reader, in which 

 the 48 schools were counted. Logistically, it was im- 

 practical for readers to follow the same sequence in 

 working with the photo-passes. 



For the repeated-counts experiment, a known sample 

 of six photographed dolphin schools (henceforth referred 

 to as the "experiment schools"), which varied in school 

 size and image quality, were counted four times. Counts 

 were done once at the start of the counting period, 

 then again every 25 days during the period. Changes 

 with time (temporal trends) due to the variables of 

 "school size," "image quality," and "reader" were tested 

 by using a repeated measures analysis of variance (RM- 

 ANOVA) model from Winer (1971; p. 337). 2 Outlier 

 counts were identified by using Shapiro-Wilk's test for 

 normality at a = 0.05 (Shapiro and Wilk, 1965). After 



-Model detailed in: Gilpatrick, J. W. Jr. (19921. Using vertical aerial 

 photographs to estimate dolphin school sizes: precision and consis- 

 tency. U.S. Dep. Commer, NOAA, Natl. Mar. Fish. Serv., Southwest 

 Fish. Sci. Cent., P.O. Box 271, La Jolla, CA 92038. Admin. Rep. 

 L.J-92-35, 20 p. 



log-transformation, data met F-test requirements of 

 homoscedasticity according to Levene's test at a = 0.05 

 (computer program BMDP7D used, Dixon et al. 1988). 

 The RM-ANOVA was computed using the software pro- 

 gram SuperANOVA (Abacus Concepts, 1990). 



Results and discussion 



Scott et al. (1985) reported that dolphin school size 

 estimates derived from aerial photographs were accu- 

 rate and more precise than visual estimates. They 

 found the standard deviation of estimates (log- trans- 

 formed) averaged 6% of school size for photographic 

 estimates and 10%-30% of school size for visual esti- 

 mates. The CV for estimates (untransformed) of the 11 

 schools used in their precision analysis averaged 8.4% 

 (range: 3.7%-15.1%) indicating slightly less precision 

 when compared with estimates presented here (avg. 

 CV: 5.4%; range: 1.2%-14.6%; Table 1). The difference 

 is explained, in part, by their statistical model, which 

 accounted for variance due not only to independent 

 repetitive counts (2 to 4 per photo-pass), but also due 

 to camera types (126- and 229-mm formats) and mul- 

 tiple photo-passes (2 to 7) for a given school. In the 

 present study, variability was minimized by use of one 

 type of camera ( 126-mm format) and by including only 

 counts of the single best photo-pass for a school. 



School size estimates averaged 146 and ranged be- 

 tween 4 and 633 (Table 1). Most schools (92%) were 

 estimated with precision that resulted in a CV of less 

 than 9.0%, and precision varied little with school size 

 (Fig. 1 and Table 2). Estimate precision tended to de- 

 crease with decreased quality of the dolphin school 

 photographs (Fig. 2). PD values (listed in Table 1) plot- 

 ted for individual readers appeared normally distrib- 

 uted, indicating no between-reader bias in counts of 

 the dolphin images. 



Repeated-count data are presented in Table 3. Out- 

 lier values for experiment school number III (Table 3) 

 resulted from reader error when dolphins were missed 

 as the marked acetate was moved from the dolphin 

 low density area of the photo-pass (in this case, the 

 beginning of the photo-pass) to the high density area. 

 Alternatively, when dolphins in the high density area 

 were counted first, the precision of the estimate was 

 improved because the majority of dolphins in the school 

 were plotted and counted at the onset; this made it 

 easier to track individual dolphins on adjoining frames. 



Within-reader CV for repeated counts averaged 3.5% . 

 Reader 2, the most experienced reader, was most pre- 

 cise in repeated counts (avg. CV: 2.6%; range: 1.4— 

 3.8%) followed by reader 3 (avg. CV: 3.4%, range: 1.5- 

 5.1%), and Reader 1 (avg. CV: 4.7%, range: 2.5-7.1%). 

 The RM-ANOVA showed significant differences between 



