2.4 Simulations
A total of 90 scenarios were simulated with each combination of drone flight pattern (n = 6), animal movement pattern (n = 3), and animal speed (n = 5) iterated 10,000 times, resulting in a total of 900,000 simulations. For each simulation, the number of times the animal was captured within the image taken by the drone was recorded and the mean and standard deviation (SD) of the raw counts were reported for model replicates to compare various combinations of our variables; these also act as measures of effect size. Accuracy of the survey counts was based on the deviation from the true value (i.e., one animal; Hone, 2008). We also report the percentage of simulations that returned the correct number of animals (n = 1), omitted the animal, or had multiple counts among scenarios. We compare subsampled landscape (transect and systematic point) counts to control scenario counts using a randomly placed, stationary animal on the landscape and report differences in mean and SD of the raw counts. A full description of the simulations, following the ODD protocol (Overview, Design concepts, and Details) for agent-based models (Grimm et al. , 2020), is provided in the Appendix.
Results
Flight pattern, animal movement pattern, and animal speed all affected the count bias with flight pattern appearing to have the most influence (Fig. 2). With one animal on the landscape, the mean and standard deviation of animal counts ranged from 0.2 ± 0.7 to 3.2 ± 2.7 animals among flight patterns, from 1.1 ± 1.1 to 1.6 ± 2.1 animals among movement patterns, and 1.2 ± 1.2 to 1.5 ± 2.0 animals among animal speeds. Although flight pattern was the most influential variable determining accurate animal counts in drone surveys, combinations of various animal movement patterns and speeds also resulted in more accurate counts of the simulated animal within various flight patterns (Fig. 2 and 3).
For flight patterns, the lawnmower pattern with 0% overlap was the least biased of all animal movement types and speeds (1.1 ± 0.6 animals, Fig. 2) with comparatively high accuracy (73.2% of simulations with correct counts; Fig. 3). The next most accurate flight pattern was the lawnmower pattern with 20% overlap (63% of simulations with correct counts) followed by the lawnmower with 40% and 60% overlap (45.7% and 33.6% of simulations with correct counts, respectively; Fig. 3). Counts increased overall with lawnmower overlap percentage, averaging 1.4 ± 0.9, 1.9 ± 1.4, and 3.2 ± 2.7 animals for 20%, 40%, and 60% overlap, respectively (Fig. 2). The transect and systematic point flight patterns were the most likely to omit the animal in the drone survey (0.2 ± 0.7 and 0.4 ± 0.5 animals, respectively; Fig. 2). The transect flight pattern very rarely returned an accurate animal count across movements and speeds (Fig. 2) and mostly omitted (87.1%) the animal, as did the systematic points (63.1%; Fig. 3). ­As the transect flight pattern captured 10% of the landscape, it should have captured the animal in 10% of our simulations; however, the average count for a moving animal was 0.2 ± 0.7, indicating that animal movement influenced survey counts, especially when compared to the average of 0.1 ± 0.3 for the stationary animal transect count. Similarly, the systematic points flight pattern, with images covering 25% of the landscape, had an average count of 0.4 ± 0.5 mobile animals, compared to 0.25 ± 0.4 stationary animals.
Animal counts were most accurate for the correlated random walk (1.1 ± 1.1 animals) among drone flight patterns for almost all animal speeds (Fig. 2). Generally, the random and biased animal walks resulted in overestimated animal counts (1.6 ± 2.1 and 1.6 ± 1.9 animals, ­­­respectively), particularly when overlap increased for lawnmower patterns from 20% to 60% (Fig. 2). The correlated random walking animal resulted in the least number of multiple counts (12.0%), with 36.0% and 37.7% of simulations having multiple counts for the random and biased random walking animal, respectively (Fig. 3). Animal movement resulted in the omission of the animal in 20.3% (correlated random walk), 32.1% (biased random walk), and 33.7% (random) of simulations (Fig. 3).
Varying the speed of the animal exhibited one clear trend among variables; increasing animal speed increased the variation around counts (i.e., lowered precision) for most flight patterns and animal walks (Fig. 2). The lawnmower pattern with 60% overlap and, to a lesser extent the 40% overlap, tended to overestimate animal counts, with average counts and variability nearly doubling, as animal speed increased from 2 to 10 m/s during random and biased random walking (Fig 2). In contrast, increasing animal speed tended to decrease multiple counts for the correlated random walk (Fig. 3). Animal speed also influenced the number of correct counts in some cases, with the percentage of correct survey counts decreasing for the 0% (66.0% and 69%) and 20% (51.8% and 53.7%) overlap for the random and biased walks, respectively, but increasing for the correlated random walk for those flight patterns (84.4% and 83.5%, respectively; Fig. 3). The number of correct survey counts also increased with animal speed for systematic points when the animal had a correlated random walk (38.3% correct at 2 m/s to 84.6% correct at 10 m/s; Fig. 3).