Abstract
The use of remote sensing to monitor animal populations has greatly expanded during the last decade. Drones (i.e. Unoccupied Aircraft Systems or UAS) provide a cost- and time-efficient remote sensing option to survey animals in various landscapes and sampling conditions. However, drone-based surveys may also introduce counting errors, especially when monitoring mobile animals. Using an agent-based model simulation approach, we evaluated the error associated with counting a single animal across various drone flight patterns under three animal movement strategies (random, directional persistence, and biased towards a resource) among five animal speeds (2, 4, 6, 8, 10 m/s). Flight patterns represented increasing spatial independence (ranging from lawnmower pattern with image overlap to systematic point counts). Simulation results indicated that flight pattern was the most important variable influencing count accuracy, followed by the type of animal movement pattern, and then animal speed. A lawnmower pattern with 0% overlap produced the most accurate count of a solitary, moving animal on a landscape (average count of 1.1 ± 0.6) regardless of the animal’s movement pattern and speed. Image overlap flight patterns were more likely to result in multiple counts even when accounting for mosaicking. Based on our simulations, we recommend using a lawnmower pattern with 0% image overlap to minimize error and augment drone efficacy for animal surveys. Our work highlights the importance of understanding interactions between animal movements and drone survey design on count accuracy to inform the development of broad applications among diverse species and ecosystems.
Keywords: agent-based model, Unoccupied Aircraft System (UAS), Unmanned Aerial Vehicle (UAV), Remotely Piloted Aircraft System (RPAS), count bias, survey error
Introduction
Drones (i.e. Unoccupied Aircraft Systems or UAS) are increasingly being used for myriad ecological applications, including animal monitoring (Koh and Wich, 2012; Vermeulen et al. , 2013; Hodgson et al. , 2018), vegetation evaluation (Olsoy et al. , 2018, 2020), and nest observation (Lyons et al. , 2019; Lachman et al. , 2020). Benefits associated with using drones in animal monitoring, compared to traditional animal survey techniques, include less time and effort in the field (McMahon, Ditmer and Forester, 2022), reduced animal disturbance compared to ground surveys (Barr et al. , 2020; Krauseet al. , 2021), and greater survey accuracy (Hodgson et al. , 2018; Jones et al. , 2020). Additionally, drones can be launched over areas inaccessible for ground surveys (Junda, Greene and Bird, 2015; Wang, Shao and Yue, 2019), provide a safer alternative for ecologists compared to occupied aircraft (Sasse, 2003; Christie et al. , 2016; Hartmann, Fishlock and Leslie, 2021), and enable creation of digital repositories of high-resolution imagery from use of advanced sensor technologies (Wang, Shao and Yue, 2019). Drone use in animal monitoring continues to increase (Linchant et al. , 2015), a trend that is exemplified by the recent annual publication rate of articles investigating animal surveys using drones during the past decade (Elmoreet al. , 2023). However, drone surveys have limitations compared to traditional methods, including relatively short battery lives (Linchant et al. , 2015), large post-processing time requirements for images (Barbedo and Vieira Koenigkan, 2018), and line of sight restrictions (Chabot and Bird, 2015; Duffy et al. , 2018).
Numerous survey methods are used in conservation science for population assessments and span a spectrum of spatial independence (Silvy, 2020). Typical drone survey methods sample an area with a lawnmower (i.e., back and forth) pattern (Elmore et al. , 2023). Belt transects are less common in drone surveys, and point counts, a common technique for ground surveys, could be adapted to drone surveys using programmed flight patterns (Silvy, 2020). Lawnmower patterns in drone surveys typically include 60-80% frontal and side overlapping of adjacent images (Fig. 1a-1d; Ezat, Fritsch and Downs, 2018; Lyons et al. , 2019; Aubertet al. , 2021). While overlapping images are necessary for mapping orthorectified landscapes (Koh and Wich, 2012), image overlap for animal monitoring can increase sampling bias due to risk of repeatedly counting individuals (Brack, Kindel and Oliveira, 2018; Lenzi et al. , 2023). Yet, common default flight settings among commercially-available drone software use overlapping lawnmower flight patterns (Harriset al. , 2019; Frazier and Singh, 2021), an approach that may not support accurate surveys.
Animal movements have the potential to influence counting accuracy in drone surveys through omission of individuals or multiple counts often caused by the same animal(s) occurring in several overlapping images (Brack, Kindel and Oliveira, 2018). Lenzi et al. (2023) mentioned “ghost” animals produced when overlapping drone images were mosaicked. These were individuals that moved during subsequent image capture, creating blurred or transparent animals on the final mosaicked photo, leading to possible erroneous counts. However, even when transect and image overlaps do not occur, multiple counts of mobile animals in drone surveys can happen (Witczuk et al. , 2018). The distance travelled by animals within a given period depends on many factors, including life history needs and a variety of abiotic (e.g., seasonal resources) and biotic (e.g., conspecific competition) influences (Nathan et al. , 2008). For example, breeding colonies of nesting shorebirds often remain on their nests (i.e., fixed locations) for long periods of time during breeding seasons (Hodgson et al. , 2016; Jones et al. , 2020). In contrast, male white-tailed deer (Odocoileus virginianus ) travel great diurnal distances (< 7.5 km) during the breeding season (Webb et al. , 2010), and wild pigs (Sus scrofa ) travel between up to 2.1 km daily, with evidence that lone boars travel longer distances than sounders (Kay et al. , 2017). Animals also exhibit changes in activity period throughout the day, with white-tailed deer (Webb et al. , 2010; Massé and Côté, 2013), black bears (Lewis and Rachlow, 2011) and wolves (Merrill and Mech, 2003) all moving more frequently during crepuscular periods. Thus, variation in animal movement patterns and speeds depend on the species ecology and current environment.
Movement models can be used to depict various animal movement patterns along a spectrum of speeds with (1) random walks representing animals dispersing randomly on the landscape, (2) correlated random walks depicting animals moving with directional persistence, mimicking something analogous to migration, and (3) biased random walks depicting animal home ranging behavior in some cases (Codling, Plank and Benhamou, 2008). These movement models challenge the common assumption among traditional survey methods of animals being detected in their original position (i.e., no movement) and can be applied to understand the influence of animal movement on drone-based survey count error. Only one study, to our knowledge, has quantified error associated with various drone flight patterns for monitoring a mobile animal (Hodgson, Peel and Kelly, 2017), the humpback whale, but their study has limited application to terrestrial systems.
Simulations represent an alternative and powerful approach to evaluate how animal movements can affect drone surveys. Simulations have been employed to investigate how various drone survey speeds and altitudes influence abundance and occupancy estimates (Baxter and Hamilton, 2018). The virtual environment can also provide insights not possible in real-world settings due to field inconsistency and other potential confounding variables (e.g., image processing, observer biases, varying detection rates). Agent-based modeling (ABM; also referred to as individual-based modeling) uses iterative computer simulations to incorporate real-world parameters in a controlled environment, modeling scenarios that can address targeted research questions (Chudzinskaet al. , 2021; Hoegh, van Manen and Haroldson, 2021). Here we used an ABM simulation approach (Grimm et al. , 2020) to (1) quantify error rates among six drone flight patterns and three common animal movement patterns at five different speeds and (2) provide suggestions for optimal drone flight patterns that minimize error associated with animal movement. Our ABM simulation approach permitted a robust examination of the potential influence of animal movements and drone flight patterns on survey count errors that would otherwise be difficult to replicate in field experiments. We predicted lawnmower flight patterns with overlapping images would overestimate true counts due to counting the same individual multiple times. We also predicted how subsampling methods such as belt transects and multiple single images (i.e., systematic point counts) would underestimate true counts due to a greater probability of omitting the moving animal. Finally, we predicted that an increase in animal speed and persistence in the directional movement of the animal would lead to overestimation as the animal could cross multiple images.
Materials and Methods
Drone Parameters
We examined the potential error among drone flight patterns and animal movement models (speed and movement pattern) using ABM simulations created in Python 3.9 (Van Rossum and Drake, 2009). To realistically approximate methodologies that conservation practitioners currently employ, the simulated drone sensor was programmed to approximate specifications of a 20-megapixel camera with a focal length of 6.8 mm and field of view of approximately 67 degrees. Flights were modeled at 61 m above ground level, representing one of several typical altitudes for animal monitoring using a multi-copter drone (McEvoy, Hall and McDonald, 2016; Wang, Shao and Yue, 2019) which has been shown to have zero or minimal behavioral impacts to several animal species (Barret al. , 2020; Krause et al. , 2021). This altitude and sensor combination produced a 1.28 cm ground sample distance and captured a ~ 50 m x 65 m ground footprint for each image. For simplicity, we adjusted the ground viewing window in simulations to a 60 x 60 m area, with grid cells in our simulation measuring 4 m by 4 m in dimension. The drone speed was simulated at 10 m/s to approximate a realistic platform speed for image capture and sharpness. To approximate real-world drone battery capabilities and line-of-sight considerations, surveys did not exceed a 30-min flight time (Raoult et al. , 2020).