Aerial Wildlife Image Repository for animal monitoring with drones in the age of artificial intelligence

Drones (unoccupied aircraft systems) have become effective tools for wildlife monitoring and conservation. Automated animal detection and classification using artificial intelligence (AI) can substantially reduce logistical and financial costs and improve drone surveys. However, the lack of annotated animal imagery for training AI is a critical bottleneck in achieving accurate performance of AI algorithms compared to other fields. To bridge this gap for drone imagery and help advance and standardize automated animal classification, we have created the Aerial Wildlife Image Repository (AWIR), which is a dynamic, interactive database with annotated images captured from drone platforms using visible and thermal cameras. The AWIR provides the first open-access repository for users to upload, annotate, and curate images of animals acquired from drones. The AWIR also provides annotated imagery and benchmark datasets that users can download to train AI algorithms to automatically detect and classify animals, and compare algorithm performance. The AWIR contains 6587 animal objects in 1325 visible and thermal drone images of predominantly large birds and mammals of 13 species in open areas of North America. As contributors increase the taxonomic and geographic diversity of available images, the AWIR will open future avenues for AI research to improve animal surveys using drones for conservation applications. Database URL: https://projectportal.gri.msstate.edu/awir/.

Man versus machine: cost and carbon emission savings of 4G-connected Artificial Intelligence technology for classifying species in camera trap images

Timely and accurate detection and identification of species are crucial for monitoring wildlife for conservation and management. Technological advances, including connectivity of camera traps to mobile phone networks and artificial intelligence (AI) algorithms for automated species identification, can potentially improve the timeliness and accuracy of species detection and identification. Adoption of this new technology, however, is often seen as cost-prohibitive as it has been difficult to calculate the cost savings or qualitative benefits over the life of the program. We developed a decision tool to quantify potential cost savings associated with incorporating the use of mobile phone network connectivity and AI technologies into monitoring programs. Using a feral cat eradication program as a case study, we used our decision tool to quantify technology-related savings in costs and carbon emissions, and compared the accuracy of AI species identification to that of experienced human observers. Over the life of the program, AI technology yielded cost savings of $0.27 M and when coupled with mobile phone network connectivity, AI saved $2.15 M and 115,838 kg in carbon emissions, with AI algorithms outperforming human observers in both speed and accuracy. Our case study demonstrates how advanced technologies can improve accuracy and cost-effectiveness and improve monitoring program efficiencies.

Automated Detection and Counting of Wild Boar in Camera Trap Images

Camera traps are becoming widely used for wildlife monitoring and management. However, manual analysis of the resulting image sets is labor-intensive, time-consuming and costly. This study shows that automated computer vision techniques can be extremely helpful in this regard, as they can rapidly and automatically extract valuable information from the images. Specific training with a set of 1600 images obtained from a study where wild animals approaching wild boar carcasses were monitored enabled the model to detect five different classes of animals automatically in their natural environment with a mean average precision of 98.11%, namely 'wild boar', 'fox', 'raccoon dog', 'deer' and 'bird'. In addition, sequences of images were automatically analyzed and the number of wild boar visits and respective group sizes were determined. This study may help to improve and speed up the monitoring of the potential spread of African swine fever virus in areas where wild boar are affected.

Deer activity levels and patterns vary along gradients of food availability and anthropogenic development

Animal activity reflects behavioral decisions that depend upon environmental context. Prior studies typically estimated activity distributions within few areas, which has limited quantitative assessment of activity changes across environmental gradients. We examined relationships between two response variables, activity level (fraction of each day spent active) and pattern (distribution of activity across a diel cycle) of white-tailed deer (Odocoileus virginianus), with four predictors-deer density, anthropogenic development, and food availability from woody twigs and agriculture. We estimated activity levels and patterns with cameras in 48 different 10.36-km2 landscapes across three larger regions. Activity levels increased with greater building density, likely due to heightened anthropogenic disturbance, but did not vary with food availability. In contrast, activity patterns responded to an interaction between twigs and agriculture, consistent with a functional response in habitat use. When agricultural land was limited, greater woody twig density was associated with reduced activity during night and evening. When agricultural land was plentiful, greater woody twig density was associated with more pronounced activity during night and evening. The region with the highest activity level also experienced the most deer-vehicle collisions. We highlight how studies of spatial variation in activity expand ecological insights on context-dependent constraints that affect wildlife behavior.

The method matters. A comparative study of biologging and camera traps as data sources with which to describe wildlife habitat selection

Habitat use is a virtually universal activity among animals and is highly relevant as regards designing wildlife management and conservation actions. This has led to the development of a great variety of methods to study it, of which resource selection functions combined with biologging-derived data (RSF) is the most widely used for this purpose. However this approach has some constraints, such as its invasiveness and high costs. Analytical approaches taking into consideration imperfect detection coupled with camera trap data (IDM) have, therefore, emerged as a non-invasive cost-effective alternative. However, despite the fact that both approaches (RSF and IDM) have been used in habitat selection studies, they should also be comparatively assessed. The objective of this work is consequently to assess them from two perspectives: explanatory and predictive. This has been done by analyzing data obtained from camera traps (60 sampling sites) and biologging (17 animals monitored: 7 red deer Cervus elaphus, 6 fallow deer Dama dama and 4 wild boar Sus scrofa) in the same periods using IDM and RSF, respectively, in Doñana National Park (southern Spain) in order to explain and predict habitat use patterns for three studied species. Our results showed discrepancies between the two approaches, as they identified different predictors as being the most relevant to determine species intensity of use, and they predicted spatial patterns of habitat use with a contrasted level of concordance, depending on species and scale. Given these results and the characteristics of each approach, we suggested that although partly comparable interpretations can be obtained with both approaches, they are not equivalent but rather complementary. The combination of data from biologging and camera traps would, therefore, appear to be suitable for the development of an analytical framework with which to describe and characterise the habitat use processes of wildlife.

Simulation-based assessment of the performance of hierarchical abundance estimators for camera trap surveys of unmarked species

Knowledge on animal abundances is essential in ecology, but is complicated by low detectability of many species. This has led to a widespread use of hierarchical models (HMs) for species abundance, which are also commonly applied in the context of nature areas studied by camera traps (CTs). However, the best choice among these models is unclear, particularly based on how they perform in the face of complicating features of realistic populations, including: movements relative to sites, multiple detections of unmarked individuals within a single survey, and low detectability. We conducted a simulation-based comparison of three HMs (Royle-Nichols, binomial N-mixture and Poisson N-mixture model) by generating groups of unmarked individuals moving according to a bivariate Ornstein-Uhlenbeck process, monitored by CTs. Under a range of simulated scenarios, none of the HMs consistently yielded accurate abundances. Yet, the Poisson N-mixture model performed well when animals did move across sites, despite accidental double counting of individuals. Absolute abundances were better captured by Royle-Nichols and Poisson N-mixture models, while a binomial N-mixture model better estimated the actual number of individuals that used a site. The best performance of all HMs was observed when estimating relative trends in abundance, which were captured with similar accuracy across these models.

Camera trapping expands the view into global biodiversity and its change

Growing threats to biodiversity demand timely, detailed information on species occurrence, diversity and abundance at large scales. Camera traps (CTs), combined with computer vision models, provide an efficient method to survey species of certain taxa with high spatio-temporal resolution. We test the potential of CTs to close biodiversity knowledge gaps by comparing CT records of terrestrial mammals and birds from the recently released Wildlife Insights platform to publicly available occurrences from many observation types in the Global Biodiversity Information Facility. In locations with CTs, we found they sampled a greater number of days (mean = 133 versus 57 days) and documented additional species (mean increase of 1% of expected mammals). For species with CT data, we found CTs provided novel documentation of their ranges (93% of mammals and 48% of birds). Countries with the largest boost in data coverage were in the historically underrepresented southern hemisphere. Although embargoes increase data providers' willingness to share data, they cause a lag in data availability. Our work shows that the continued collection and mobilization of CT data, especially when combined with data sharing that supports attribution and privacy, has the potential to offer a critical lens into biodiversity. This article is part of the theme issue 'Detecting and attributing the causes of biodiversity change: needs, gaps and solutions'.

Beaver Dams and Fallen Trees as Ecological Corridors Allowing Movements of Mammals across Water Barriers-A Case Study with the Application of Novel Substrate for Tracking Tunnels

Physical obstacles within animal habitats create barriers to individual movements. To cross those barriers, specific corridors are used, some of them created by keystone species such as Eurasian beavers (Castor fiber). Their dams on rivers may also increase habitat connectivity for terrestrial mammals, but the significance of that function has never been quantified. To investigate this, we placed tracking tunnels on beaver dams, fallen trees, and-as a control-on floating rafts. Additionally, we tested kinetic sand as a novel substrate for collecting tracks and found the paws of small mustelids precisely imprinted in that medium, allowing easy identification. However, we needed to lump all shrews and rodents smaller than water voles (Arvicola amphibius) into one category as they can only be detected but not identified. The highest mammalian activity was observed on dams, as they may provide shelter, offering protection from predators during a river crossing or permanent residence, and even the opportunity to hunt invertebrates. Slightly higher diversity was found on logs because of a higher proportion of mustelids, which select exposed locations for scent marking. Our results increase our body of knowledge about the beaver as an ecosystem engineer and provide a novel tool for the monitoring of mammal activity.

Evidence on the efficacy of small unoccupied aircraft systems (UAS) as a survey tool for North American terrestrial, vertebrate animals: a systematic map

BACKGROUND

Small unoccupied aircraft systems (UAS) are replacing or supplementing occupied aircraft and ground-based surveys in animal monitoring due to improved sensors, efficiency, costs, and logistical benefits. Numerous UAS and sensors are available and have been used in various methods. However, justification for selection or methods used are not typically offered in published literature. Furthermore, existing reviews do not adequately cover past and current UAS applications for animal monitoring, nor their associated UAS/sensor characteristics and environmental considerations. We present a systematic map that collects and consolidates evidence pertaining to UAS monitoring of animals.

METHODS

We investigated the current state of knowledge on UAS applications in terrestrial animal monitoring by using an accurate, comprehensive, and repeatable systematic map approach. We searched relevant peer-reviewed and grey literature, as well as dissertations and theses, using online publication databases, Google Scholar, and by request through a professional network of collaborators and publicly available websites. We used a tiered approach to article exclusion with eligible studies being those that monitor (i.e., identify, count, estimate, etc.) terrestrial vertebrate animals. Extracted metadata concerning UAS, sensors, animals, methodology, and results were recorded in Microsoft Access. We queried and catalogued evidence in the final database to produce tables, figures, and geographic maps to accompany this full narrative review, answering our primary and secondary questions.

REVIEW FINDINGS

We found 5539 articles from our literature searches of which 216 were included with extracted metadata categories in our database and narrative review. Studies exhibited exponential growth over time but have levelled off between 2019 and 2021 and were primarily conducted in North America, Australia, and Antarctica. Each metadata category had major clusters and gaps, which are described in the narrative review.

CONCLUSIONS

Our systematic map provides a useful synthesis of current applications of UAS-animal related studies and identifies major knowledge clusters (well-represented subtopics that are amenable to full synthesis by a systematic review) and gaps (unreported or underrepresented topics that warrant additional primary research) that guide future research directions and UAS applications. The literature for the use of UAS to conduct animal surveys has expanded intensely since its inception in 2006 but is still in its infancy. Since 2015, technological improvements and subsequent cost reductions facilitated widespread research, often to validate UAS technology to survey single species with application of descriptive statistics over limited spatial and temporal scales. Studies since the 2015 expansion have still generally focused on large birds or mammals in open landscapes of 4 countries, but regulations, such as maximum altitude and line-of-sight limitations, remain barriers to improved animal surveys with UAS. Critical knowledge gaps include the lack of (1) best practices for using UAS to conduct standardized surveys in general, (2) best practices to survey whole wildlife communities in delineated areas, and (3) data on factors affecting bias in counting animals from UAS images. Promising advances include the use of thermal sensors in forested environments or nocturnal surveys and the development of automated or semi-automated machine-learning algorithms to accurately detect, identify, and count animals from UAS images.

Camera-trapping estimates of the relative population density of Sympetrum dragonflies: application to multihabitat users in agricultural landscapes

Although camera trapping has been effectively used for wildlife monitoring, its application to multihabitat insects (i.e., insects requiring terrestrial and aquatic ecosystems) is limited. Among such insects, perching dragonflies of the genus Sympetrum (darter dragonflies) are agroenvironmental indicators that substantially contribute to agricultural biodiversity. To examine whether custom-developed camera traps for perching dragonflies can be used to assess the relative population density of darter dragonflies, camera trapping, a line-transect survey of mature adult dragonflies, and a line-transect survey of exuviae were conducted for three years in rice paddy fields in Japan. The detection frequency of camera traps in autumn was significantly correlated with the density index of mature adults recorded during the transect surveys in the same season for both Sympetrum infuscatum and other darter species. In analyses of camera-detection frequency in autumn and exuviae in early summer, a significant correlation was observed between the camera-detection frequency of mature adults and the exuviae-density index in the following year for S. infuscatum; however, a similar correlation was not observed for other darter species. These results suggest that terrestrial camera trapping has the potential to be effective for monitoring the relative density of multihabitat users such as S. infuscatum, which shows frequent perching behavior and relatively short-distance dispersal.

The inappropriate use of time-to-independence biases estimates of activity patterns of free-ranging mammals derived from camera traps

Measuring and comparing activity patterns provide key insights into the behavioral trade-offs that result in animal activity and their extrinsic and intrinsic drivers. Camera traps are a recently emerged source of data for sampling animal activity used to estimate activity patterns. However, nearly 70% of studies using such data to estimate activity patterns apply a time-to-independence data filter to discard appreciable periods of sampling effort. This treatment of activity as a discrete event emerged from the use of camera trap data to estimate animal abundances, but does not reflect the continuous nature of behavior, and may bias resulting estimates of activity patterns. We used a large, freely available camera trap dataset to test the effects of time to independence on the estimated activity of eight medium- to large-sized African mammals. We show that discarding data through the use of time-to-independence filters causes substantial losses in sample sizes and differences in the estimated activity of species. Activity patterns estimated for herbivore species were more affected by the application of time-to-independence data filters than carnivores, this extending to estimates of potential interactions (activity overlap) between herbivore species. We hypothesize that this pattern could reflect the typically more abundant, social, and patch-specific foraging patterns of herbivores and suggest that this effect may bias estimates of predator-prey interactions. Activity estimates of rare species, with less data available, may be particularly vulnerable to loss of data through the application of time-to-independence data filters. We conclude that the application of time-to-independence data filters in camera trap-based estimates of activity patterns is not valid and should not be used.

Deer Behavior Affects Density Estimates With Camera Traps, but Is Outweighed by Spatial Variability

Density is a key trait of populations and an essential parameter in ecological research, wildlife conservation and management. Several models have been developed to estimate population density based on camera trapping data, including the random encounter model (REM) and camera trap distance sampling (CTDS). Both models need to account for variation in animal behavior that depends, for example, on the species and sex of the animals along with temporally varying environmental factors. We examined whether the density estimates of REM and CTDS can be improved for Europe’s most numerous deer species, by adjusting the behavior-related model parameters per species and accounting for differences in movement speeds between sexes, seasons, and years. Our results showed that bias through inadequate consideration of animal behavior was exceeded by the uncertainty of the density estimates, which was mainly influenced by variation in the number of independent observations between camera trap locations. The neglection of seasonal and annual differences in movement speed estimates for REM overestimated densities of red deer in autumn and spring by ca. 14%. This GPS telemetry-derived parameter was found to be most problematic for roe deer females in summer and spring when movement behavior was characterized by small-scale displacements relative to the intervals of the GPS fixes. In CTDS, density estimates of red deer improved foremost through the consideration of behavioral reactions to the camera traps (avoiding bias of max. 19%), while species-specific delays between photos had a larger effect for roe deer. In general, the applicability of both REM and CTDS would profit profoundly from improvements in their precision along with the reduction in bias achieved by exploiting the available information on animal behavior in the camera trap data.

Camera traps reveal overlap and seasonal variation in the diel activity of arboreal and semi-arboreal mammals

Our study aimed to investigate seasonal variation in the activity of arboreal and semi-arboreal mammals and investigate their overlap in temporal activity, as well temporal shifts in activity because of behavioural interference. In our camera trapping study in a fragmented landscape in south-eastern Australia, a total of ten arboreal and semi-arboreal species were found, with 35,671 independent observations recorded over 6517 camera trap nights. All species were found to be nocturnal; however, a notable number of daytime observations were made for several species (i.e. brown antechinus, Antechinus stuartii; sugar glider, Petaurus breviceps; bush rat, Rattus fuscipes; brown rat, Rattus norvegicus). Seasonal variations in diel activity were observed through an increase in crepuscular activity in spring and summer for antechinus, sugar gliders, brown rats, brushtail possums, Trichosurus vulpecula and ringtail possums, Pseudocheirus peregrinus. Diel activity overlap between species was high, that is 26/28 species comparisons had overlap coefficients (Δ) > 0.75. The species pair with the least amount of overlap was between southern bobucks, Trichosurus cunninghami and brown antechinus (Δ4 = 0.66). The species pair with the most overlap was between the native sugar glider and introduced brown rat (Δ4 = 0.93). When comparing the activity of sugar gliders in sites with low and high abundance of brown rats, sugar gliders appear to shift their activity relative to the brown rats. Similarly, behavioural interference was also observed between antechinus and sugar gliders, and when comparing sites of low and high abundance of sugar glider, antechinus had a shift in activity. Our work provides some of the first quantification of temporal patterns for several of the species in this study, and the first for a community of arboreal and semi-arboreal mammals. Our results indicate that some shifts in behaviour are potentially occurring in response to behavioural interference, allowing for coexistence by means of temporal partitioning.