2.2 Landscape and Drone Flight Patterns
We simulated six drone flight patterns, which increased along a spectrum of spatial independence among images and included the following commonly used flight patterns: (1) a lawnmower pattern with 60% image overlap (Fig. 1a), (2) a lawnmower pattern with 40% image overlap (Fig. 1b), (3) a lawnmower pattern with 20% image overlap (Fig. 1c), (4) a lawnmower pattern with 0% image overlap where images touched (Fig. 1d), (5) a randomized belt transect (Fig. 1e), and (6) systematic points (Fig. 1f). To ensure the assumption that the animal was 100% available and detectable during the simulated survey, the landscape dimensions were slightly revised for the lawnmower patterns with 20, 40, and 60% image overlap to ensure complete coverage by the drone imagery. The lawnmower patterns with 20% and 40% image overlap covered a 242,064 m2 (492 x 492 m) landscape; whereas for 60% overlap, the landscape size was adjusted to 219,024 m2 (468 x 468 m). For the lawnmower pattern with 0% image overlap, transect, and systematic point flight patterns, the landscape size was fixed at 230,400 m2 (480 x 480 m).
Transect surveys included one horizontal belt transect with a length of 384 m (80% of the total landscape length) and a width of 60 m (image width; Fig. 1e). Image captures from transects were programmed to have 60% frontal overlap, capturing imagery of 10% of the total landscape. Transects were generated to include stochasticity among simulations by randomly selecting the initial x and y coordinates for each replicate in places that would allow the entire transect to be placed horizontally across the landscape. The systematic points flight pattern simulated 16 image captures evenly distributed across the landscape (Fig. 1f), which amounts to the same number of images captured by the transect survey. However, since the systematic points flight pattern did not exhibit any image overlap it was able to capture 25% of the total landscape. The animal was counted when it was located inside the image viewing window. To account for approaches where multiple images would be stitched into an orthomosaic (Frazier and Singh, 2021), an animal was not counted in an image if it had not moved more than 4 m from its previous location as the animal would have remained within the same grid cell. Previous studies describing “ghost” animal issues (Brack, Kindel and Oliveira, 2018; Lenzi et al. , 2023) do not detail how far animals moved when creating discrepancies, but in our case movements greater than 4 m were assumed to be large enough to cause issues with post-processing software within the simulations.