Counting Mixed Breeding Aggregations of Animal Species Using Drones: Lessons from Waterbirds on Semi-Automation

Abstract

Using drones to count wildlife saves time and resources and allows access to difficult or dangerous areas. We collected drone imagery of breeding waterbirds at colonies in the Okavango Delta (Botswana) and Lowbidgee floodplain (Australia). We developed a semi-automated counting method, using machine learning, and compared effectiveness of freeware and payware in identifying and counting waterbird species (targets) in the Okavango Delta. We tested transferability to the Australian breeding colony. Our detection accuracy (targets), between the training and test data, was 91% for the Okavango Delta colony and 98% for the Lowbidgee floodplain colony. These estimates were within 1-5%, whether using freeware or payware for the different colonies. Our semi-automated method was 26% quicker, including development, and 500% quicker without development, than manual counting. Drone data of waterbird colonies can be collected quickly, allowing later counting with minimal disturbance. Our semi-automated methods efficiently provided accurate estimates of nesting species of waterbirds, even with complex backgrounds. This could be used to track breeding waterbird populations around the world, indicators of river and wetland health, with general applicability for monitoring other taxa.Remote Sens. 2020, 12, 1185 2 of 17 semi-automated methods. The former can be extremely labour-intensive and consequently expensive, particularly for large aggregations of wildlife [26], further complicated when more than one species is counted. Semi-automated methods, including the counting of animals from photographs (e.g., camera traps) and drone imagery, are increasingly being developed around the world [27]. These methods reduce the time required to count and process drone images [28], accelerating the data entry stage and encouraging the use of drones as scientific tools for management. Such benefits allow for real-time monitoring and management decisions and could, for example, assist in the targeted delivery of environmental flows for waterbird breeding events [29].Generally, semi-automated counting methods are most effective for species where there are strong contrasts against the backgrounds, particularly when background colours and shapes are consistent [28]. They can distinguish large single species aggregations on relatively simple backgrounds [30][31][32], up to sixteen avian species (numbering in the hundreds) on simple single colour backgrounds, such as oceans [33,34], or single species aggregations of hundreds of thousands on complex backgrounds [3].Development of flexible, repeatable and efficient methods, using open source software, is important in ensuring methods are applicable across a range of datasets [35,36]. Further, there are potential cost implications of processing data, given that some processing software can be expensive (i.e., compulsory licence fees, called 'payware' in this paper) and so are often only accessible to large organisations in high-income countries [37]. Open source software, or software with optional lic...

Automated detection of wildlife using drones: Synthesis, opportunities and constraints

“…Nine unique survey campaigns used RGB sensors to collect images, while four used infrared cameras and two used and compared both sensor types (Table 2). While there was a considerable overlap in the size of species surveyed with both types of sensor, only RGB sensors were used to survey species with an average body size of less than 4 kg which included several species of waterbirds with the smallest being decoys that simulated Greated Crested Tern weighing 0.34 kg (Afan et al, 2018;Francis et al, 2020;Hodgson et al, 2018;Hong et al, 2019;Liu et al, 2015;Rush et al, 2018) (Table 1).…”

“…Birds were only surveyed with RGB sensors, while koalas were the only arboreal mammal surveyed and this was conducted with an infrared sensor (Table 1). This is likely due to the bird species surveyed generally occurring in relatively homogenous habitat with no canopy cover (Afan et al, 2018;Francis et al, 2020;Hodgson et al, 2018;Hong et al, 2019;Rush et al, 2018), meaning they were easily detectable using the visible spectrum, compared to the koalas surveyed by Corcoran et al (2019) which tended to be occluded by canopy. In infrared imaging, the heat given off by the koalas was visible even in cases where individuals were partially covered by branches, therefore increasing probability of detection in more heterogeneous habitats with dense vegetation (Corcoran et al, 2019).…”

“…The explicit impacts of canopy cover and topography on automated detection remain largely unknown, although a few recent studies suggest it is feasible to overcome the challenges presented by more complex survey areas with the appropriate combination of drone platform, sensor and detection method (Corcoran et al, 2019;Oishi et al, 2018). There is also little known about how differences in platforms, sensors and design of drone wildlife surveys affect the number of false positives yielded by automated detection methods, as few studies to date have explicitly reported precision rates despite many stating misidentified objects were a challenge to producing accurate counts of the target species (Corcoran et al, 2019;Francis et al, 2020;Gorkin et al, 2020;Hong et al, 2019;Kellenberger, Marcos, Courty, et al, 2018;Kellenberger et al, 2017Kellenberger et al, , 2019Rey et al, 2017). Further studies should be carried out to investigate the effect of these variables on the probability of detection and precision rate for species of different sizes in both RGB and infrared imagery to better inform the selection of suitable sites for automated detection and counting of vulnerable wildlife species to provide insight for conservation managers (Anderson & Gaston, 2013;Burke et al, 2019;McCafferty et al, 2015;Renwick et al, 2011).…”

Accuracy and precision of citizen scientist animal counts from drone imagery

“…The accuracy we report may be higher than observed in other mammal or bird species because pinnipeds’ large size and distinctive shapes make them excellent subjects [ 42 ]. Additionally, a more expensive drone with a higher resolution camera could be used to increase accuracy; however, these tend to be larger drones with more potential for animal disturbance [ 43 ].…”

Drones, Gulls and Urbanity: Interaction between New Technologies and Human Subsidized Species in Coastal Areas

“…The breeding season is a short-term process but intense in most seabirds, whereas the non-breeding period is a longer lapse when birds spread outside their central breeding territories. Available research using drones in waterbird ecology have been conducted mostly during breeding stages, especially those focusing on colonial seabirds [20,21], probably in response to the complexity of conducting research after individual dispersion during post-breeding. In this regard, responses to drones have been evaluated in colonial seabirds only during breeding [17,22], but information about the reactions of aggressive seabirds as gulls against drones are scarce during breeding [18] and unknown during non-breeding seasons, respectively.…”