Automatically identifying, counting, and describing wild animals in camera-trap images with deep learning.

Having accurate, detailed, and up-to-date information about the location and behavior of animals in the wild would improve our ability to study and conserve ecosystems. We investigate the ability to automatically, accurately, and inexpensively collect such data, which could help catalyze the transformation of many fields of ecology, wildlife biology, zoology, conservation biology, and animal behavior into "big data" sciences. Motion-sensor "camera traps" enable collecting wildlife pictures inexpensively, unobtrusively, and frequently. However, extracting information from these pictures remains an expensive, time-consuming, manual task. We demonstrate that such information can be automatically extracted by deep learning, a cutting-edge type of artificial intelligence. We train deep convolutional neural networks to identify, count, and describe the behaviors of 48 species in the 3.2 million-image Snapshot Serengeti dataset. Our deep neural networks automatically identify animals with >93.8% accuracy, and we expect that number to improve rapidly in years to come. More importantly, if our system classifies only images it is confident about, our system can automate animal identification for 99.3% of the data while still performing at the same 96.6% accuracy as that of crowdsourced teams of human volunteers, saving >8.4 y (i.e., >17,000 h at 40 h/wk) of human labeling effort on this 3.2 million-image dataset. Those efficiency gains highlight the importance of using deep neural networks to automate data extraction from camera-trap images, reducing a roadblock for this widely used technology. Our results suggest that deep learning could enable the inexpensive, unobtrusive, high-volume, and even real-time collection of a wealth of information about vast numbers of animals in the wild.

A generalized approach for producing, quantifying, and validating citizen science data from wildlife images.

Citizen science has the potential to expand the scope and scale of research in ecology and conservation, but many professional researchers remain skeptical of data produced by nonexperts. We devised an approach for producing accurate, reliable data from untrained, nonexpert volunteers. On the citizen science website www.snapshotserengeti.org, more than 28,000 volunteers classified 1.51 million images taken in a large-scale camera-trap survey in Serengeti National Park, Tanzania. Each image was circulated to, on average, 27 volunteers, and their classifications were aggregated using a simple plurality algorithm. We validated the aggregated answers against a data set of 3829 images verified by experts and calculated 3 certainty metrics-level of agreement among classifications (evenness), fraction of classifications supporting the aggregated answer (fraction support), and fraction of classifiers who reported "nothing here" for an image that was ultimately classified as containing an animal (fraction blank)-to measure confidence that an aggregated answer was correct. Overall, aggregated volunteer answers agreed with the expert-verified data on 98% of images, but accuracy differed by species commonness such that rare species had higher rates of false positives and false negatives. Easily calculated analysis of variance and post-hoc Tukey tests indicated that the certainty metrics were significant indicators of whether each image was correctly classified or classifiable. Thus, the certainty metrics can be used to identify images for expert review. Bootstrapping analyses further indicated that 90% of images were correctly classified with just 5 volunteers per image. Species classifications based on the plurality vote of multiple citizen scientists can provide a reliable foundation for large-scale monitoring of African wildlife.

Applying a random encounter model to estimate lion density from camera traps in Serengeti National Park, Tanzania.

The random encounter model (REM) is a novel method for estimating animal density from camera trap data without the need for individual recognition. It has never been used to estimate the density of large carnivore species, despite these being the focus of most camera trap studies worldwide. In this context, we applied the REM to estimate the density of female lions (Panthera leo) from camera traps implemented in Serengeti National Park, Tanzania, comparing estimates to reference values derived from pride census data. More specifically, we attempted to account for bias resulting from non-random camera placement at lion resting sites under isolated trees by comparing estimates derived from night versus day photographs, between dry and wet seasons, and between habitats that differ in their amount of tree cover. Overall, we recorded 169 and 163 independent photographic events of female lions from 7,608 and 12,137 camera trap days carried out in the dry season of 2010 and the wet season of 2011, respectively. Although all REM models considered over-estimated female lion density, models that considered only night-time events resulted in estimates that were much less biased relative to those based on all photographic events. We conclude that restricting REM estimation to periods and habitats in which animal movement is more likely to be random with respect to cameras can help reduce bias in estimates of density for female Serengeti lions. We highlight that accurate REM estimates will nonetheless be dependent on reliable measures of average speed of animal movement and camera detection zone dimensions. © 2015 The Authors. Journal of Wildlife Management published by Wiley Periodicals, Inc. on behalf of The Wildlife Society.

Burkholderia pseudomallei isolates in 2 pet iguanas, California, USA.

Burkholderia pseudomallei, the causative agent of melioidosis, was isolated from abscesses of 2 pet green iguanas in California, USA. The international trade in iguanas may contribute to importation of this pathogen into countries where it is not endemic and put persons exposed to these animals at risk for infection.

Cheetahs and wild dogs show contrasting patterns of suppression by lions.

Top predators can dramatically suppress populations of smaller predators, with cascading effects throughout communities, and this pressure is often unquestioningly accepted as a constraint on mesopredator populations. In this study, we reassess whether African lions suppress populations of cheetahs and African wild dogs and examine possible mechanisms for coexistence between these species. Using long-term records from Serengeti National Park, we tested 30 years of population data for evidence of mesopredator suppression, and we examined six years of concurrent radio-telemetry data for evidence of large-scale spatial displacement. The Serengeti lion population nearly tripled between 1966 and 1998; during this time, wild dogs declined but cheetah numbers remained largely unchanged. Prior to their local extinction, wild dogs primarily occupied low lion density areas and apparently abandoned the long-term study area as the lion population 'saturated' the region. In contrast, cheetahs mostly utilized areas of high lion density, and the stability of the cheetah population indicates that neither high levels of lion-inflicted mortality nor behavioural avoidance inflict sufficient demographic consequences to translate into population-level effects. Population data from fenced reserves in southern Africa revealed a similar contrast between wild dogs and cheetahs in their ability to coexist with lions. These findings demonstrate differential responses of subordinate species within the same guild and challenge a widespread perception that lions undermine cheetah conservation efforts. Paired with several recent studies that document fine-scale lion-avoidance by cheetahs, this study further highlights fine-scale spatial avoidance as a possible mechanism for mitigating mesopredator suppression.

α-Synuclein binds the K(ATP) channel at insulin-secretory granules and inhibits insulin secretion.

α-Synuclein has been studied in numerous cell types often associated with secretory processes. In pancreatic ß-cells, α-synuclein might therefore play a similar role by interacting with organelles involved in insulin secretion. We tested for α-synuclein localizing to insulin-secretory granules and characterized its role in glucose-stimulated insulin secretion. Immunohistochemistry and fluorescent sulfonylureas were used to test for α-synuclein localization to insulin granules in ß-cells, immunoprecipitation with Western blot analysis for interaction between α-synuclein and K(ATP) channels, and ELISA assays for the effect of altering α-synuclein expression up or down on insulin secretion in INS1 cells or mouse islets, respectively. Differences in cellular phenotype between α-synuclein knockout and wild-type ß-cells were found by using confocal microscopy to image the fluorescent insulin biosensor Ins-C-emGFP and by using transmission electron microscopy. The results show that anti-α-synuclein antibodies labeled secretory organelles within ß-cells. Anti-α-synuclein antibodies colocalized with K(ATP) channel, anti-insulin, and anti-C-peptide antibodies. α-Synuclein coimmunoprecipitated in complexes with K(ATP) channels. Expression of α-synuclein downregulated insulin secretion at 2.8 mM glucose with little effect following 16.7 mM glucose stimulation. α-Synuclein knockout islets upregulated insulin secretion at 2.8 and 8.4 mM but not 16.7 mM glucose, consistent with the depleted insulin granule density at the ß-cell surface membranes observed in these islets. These findings demonstrate that α-synuclein interacts with K(ATP) channels and insulin-secretory granules and functionally acts as a brake on secretion that glucose stimulation can override. α-Synuclein might play similar roles in diabetes as it does in other degenerative diseases, including Alzheimer's and Parkinson's diseases.

QUANTIFYING THE SEVERITY OF GIRAFFE SKIN DISEASE VIA PHOTOGRAMMETRY ANALYSIS OF CAMERA TRAP DATA.

Developing techniques to quantify the spread and severity of diseases afflicting wildlife populations is important for disease ecology, animal ecology, and conservation. Giraffes (Giraffa camelopardalis) are in the midst of a dramatic decline, but it is not known whether disease is playing an important role in the broad-scale population reductions. A skin disorder referred to as giraffe skin disease (GSD) was recorded in 1995 in one giraffe population in Uganda. Since then, GSD has been detected in 13 populations in seven African countries, but good descriptions of the severity of this disease are not available. We photogrammetrically analyzed camera trap images from both Ruaha and Serengeti National parks in Tanzania to quantify GSD severity. Giraffe skin disease afflicts the limbs of giraffes in Tanzania, and we quantified severity by measuring the vertical length of the GSD lesion in relation to the total leg length. Applying the Jenks natural breaks algorithm to the lesion proportions that we derived, we classified individual giraffes into disease categories (none, mild, moderate, and severe). Scaling up to the population level, we predicted the proportion of the Ruaha and Serengeti giraffe populations with mild, moderate, and severe GSD. This study serves to demonstrate that camera traps presented an informative platform for examinations of skin disease ecology.

In the absence of a "landscape of fear": How lions, hyenas, and cheetahs coexist.

Aggression by top predators can create a "landscape of fear" in which subordinate predators restrict their activity to low-risk areas or times of day. At large spatial or temporal scales, this can result in the costly loss of access to resources. However, fine-scale reactive avoidance may minimize the risk of aggressive encounters for subordinate predators while maintaining access to resources, thereby providing a mechanism for coexistence. We investigated fine-scale spatiotemporal avoidance in a guild of African predators characterized by intense interference competition. Vulnerable to food stealing and direct killing, cheetahs are expected to avoid both larger predators; hyenas are expected to avoid lions. We deployed a grid of 225 camera traps across 1,125 km2 in Serengeti National Park, Tanzania, to evaluate concurrent patterns of habitat use by lions, hyenas, cheetahs, and their primary prey. We used hurdle models to evaluate whether smaller species avoided areas preferred by larger species, and we used time-to-event models to evaluate fine-scale temporal avoidance in the hours immediately surrounding top predator activity. We found no evidence of long-term displacement of subordinate species, even at fine spatial scales. Instead, hyenas and cheetahs were positively associated with lions except in areas with exceptionally high lion use. Hyenas and lions appeared to actively track each, while cheetahs appear to maintain long-term access to sites with high lion use by actively avoiding those areas just in the hours immediately following lion activity. Our results suggest that cheetahs are able to use patches of preferred habitat by avoiding lions on a moment-to-moment basis. Such fine-scale temporal avoidance is likely to be less costly than long-term avoidance of preferred areas: This may help explain why cheetahs are able to coexist with lions despite high rates of lion-inflicted mortality, and highlights reactive avoidance as a general mechanism for predator coexistence.

The spatial distribution of African savannah herbivores: species associations and habitat occupancy in a landscape context.

Herbivores play an important role in determining the structure and function of tropical savannahs. Here, we (i) outline a framework for how interactions among large mammalian herbivores, carnivores and environmental variation influence herbivore habitat occupancy in tropical savannahs. We then (ii) use a Bayesian hierarchical model to analyse camera trap data to quantify spatial patterns of habitat occupancy for lions and eight common ungulates of varying body size across an approximately 1100 km(2) landscape in the Serengeti ecosystem. Our results reveal strong positive associations among herbivores at the scale of the entire landscape. Lions were positively associated with migratory ungulates but negatively associated with residents. Herbivore habitat occupancy differed with body size and migratory strategy: large-bodied migrants, at less risk of predation and able to tolerate lower quality food, were associated with high NDVI, while smaller residents, constrained to higher quality forage, avoided these areas. Small herbivores were strongly associated with fires, likely due to the subsequent high-quality regrowth, while larger herbivores avoided burned areas. Body mass was strongly related to herbivore habitat use, with larger species more strongly associated with riverine and woodlands than smaller species. Large-bodied migrants displayed diffuse habitat occupancy, whereas smaller species demonstrated fine-scale occupancy reflecting use of smaller patches of high-quality habitat. Our results demonstrate the emergence of strong positive spatial associations among a diverse group of savannah herbivores, while highlighting species-specific habitat selection strongly determined by herbivore body size.This article is part of the themed issue 'Tropical grassy biomes: linking ecology, human use and conservation'.

Burden of Bites by Dogs and Other Animals in Los Angeles County, California, 2009-2011.

OBJECTIVES: The objective of this study was to analyze all animal-to-human bite reports during a 3-year period from a regional surveillance database. Results helped to inform local efforts to reduce and prevent animal-to-human bites. METHODS: We reviewed all cases of animal-to-human bites occurring from 2009 through 2011 that were reported to the Los Angeles County Department of Public Health's Animal Bites Database. We collected data on the bite victim's date of birth, age, and address; bite circumstances (ie, date, time, location, how bite occurred); anatomic site and treatment of bite; type of reporting facility; and breed and management of biting animal. RESULTS: From 2009 through 2011, 26   169 animal-to-human bites were reported, of which 23   103 (88%) were dog bites. Most animal-to-human bites (n = 7673, 29%) occurred between 4 pm and 8 pm and peaked during the month of July (n = 2663, 10%). Most animal-to-human bites occurred outdoors (n = 8772, 34%) and while victims engaged in recreational activities (n = 4353, 17%). The hands were the most common injury site (n = 9130, 35%), and only 1% of animal bites (n = 267) resulted in hospitalization. A total of 4115 bite victims (16%) received tetanus vaccinations. Of all animal-to-human bite cases, medical organizations reported 13   451 (51%), and animal control agencies reported 10   682 (41%). CONCLUSIONS: Animal-to-human bites can often lead to medical complications. Surveillance is essential in helping to identify, manage, and reduce these highly preventable injuries and direct public health actions and policies on animal bite risk and prevention.

Snapshot Serengeti, high-frequency annotated camera trap images of 40 mammalian species in an African savanna.

Camera traps can be used to address large-scale questions in community ecology by providing systematic data on an array of wide-ranging species. We deployed 225 camera traps across 1,125 km(2) in Serengeti National Park, Tanzania, to evaluate spatial and temporal inter-species dynamics. The cameras have operated continuously since 2010 and had accumulated 99,241 camera-trap days and produced 1.2 million sets of pictures by 2013. Members of the general public classified the images via the citizen-science website www.snapshotserengeti.org. Multiple users viewed each image and recorded the species, number of individuals, associated behaviours, and presence of young. Over 28,000 registered users contributed 10.8 million classifications. We applied a simple algorithm to aggregate these individual classifications into a final 'consensus' dataset, yielding a final classification for each image and a measure of agreement among individual answers. The consensus classifications and raw imagery provide an unparalleled opportunity to investigate multi-species dynamics in an intact ecosystem and a valuable resource for machine-learning and computer-vision research.

Fear of darkness, the full moon and the nocturnal ecology of African lions.

Nocturnal carnivores are widely believed to have played an important role in human evolution, driving the need for night-time shelter, the control of fire and our innate fear of darkness. However, no empirical data are available on the effects of darkness on the risks of predation in humans. We performed an extensive analysis of predatory behavior across the lunar cycle on the largest dataset of lion attacks ever assembled and found that African lions are as sensitive to moonlight when hunting humans as when hunting herbivores and that lions are most dangerous to humans when the moon is faint or below the horizon. At night, people are most active between dusk and 10:00 pm, thus most lion attacks occur in the first weeks following the full moon (when the moon rises at least an hour after sunset). Consequently, the full moon is a reliable indicator of impending danger, perhaps helping to explain why the full moon has been the subject of so many myths and misconceptions.

Sport hunting, predator control and conservation of large carnivores.

Sport hunting has provided important economic incentives for conserving large predators since the early 1970's, but wildlife managers also face substantial pressure to reduce depredation. Sport hunting is an inherently risky strategy for controlling predators as carnivore populations are difficult to monitor and some species show a propensity for infanticide that is exacerbated by removing adult males. Simulation models predict population declines from even moderate levels of hunting in infanticidal species, and harvest data suggest that African countries and U.S. states with the highest intensity of sport hunting have shown the steepest population declines in African lions and cougars over the past 25 yrs. Similar effects in African leopards may have been masked by mesopredator release owing to declines in sympatric lion populations, whereas there is no evidence of overhunting in non-infanticidal populations of American black bears. Effective conservation of these animals will require new harvest strategies and improved monitoring to counter demands for predator control by livestock producers and local communities.