Search

Motion‐activated wildlife cameras (or “camera traps”) are frequently used to remotely and noninvasively observe animals. The vast number of images collected from camera trap projects has prompted some biologists to employ machine learning algorithms to automatically recognize species in these images, or at least filter‐out images that do not contain animals. These approaches are often limited by model transferability, as a model trained to recognize species from one location might not work as well for the same species in different locations. Furthermore, these methods often require advanced computational skills, making them inaccessible to many biologists. We used 3 million camera trap images from 18 studies in 10 states across the United States of America to train two deep neural networks, one that recognizes 58 species, the “species model,” and one that determines if an image is empty or if it contains an animal, the “empty‐animal model.” Our species model and empty‐animal model had accuracies of 96.8% and 97.3%, respectively. Furthermore, the models performed well on some out‐of‐sample datasets, as the species model had 91% accuracy on species from Canada (accuracy range 36%–91% across all out‐of‐sample datasets) and the empty‐animal model achieved an accuracy of 91%–94% on out‐of‐sample datasets from different continents. Our software addresses some of the limitations of using machine learning to classify images from camera traps. By including many species from several locations, our species model is potentially applicable to many camera trap studies in North America. We also found that our empty‐animal model can facilitate removal of images without animals globally. We provide the trained models in an R package (MLWIC2: Machine Learning for Wildlife Image Classification in R), which contains Shiny Applications that allow scientists with minimal programming experience to use trained models and train new models in six neural network architectures with varying depths., We provide an R package (MLWIC2), which includes two models with 97% validation accuracy, one that recognizes 58 animal species from camera trap images and another that removes empty images. The R package includes Shiny Applications, allowing users to apply trained models and train new models using a point‐and‐click interface.

Motion-activated cameras (camera traps) are increasingly used in ecological and management studies for remotely observing wildlife and have been regarded as among the most powerful tools for wildlife research. However, studies involving camera traps result in millions of images that need to be analyzed, typically by visually observing each image, in order to extract data that can be used in ecological analyses. We trained machine learning models using convolutional neural networks with the ResNet-18 architecture and 3,367,383 images to automatically classify wildlife species from camera trap images obtained from five states across the United States. We tested our model on an independent subset of images not seen during training from the United States and on an out-of-sample (or out-of-distribution in the machine learning literature) dataset of ungulate images from Canada. We also tested the ability of our model to distinguish empty images from those with animals in another out-of-sample dataset from Tanzania, containing a faunal community that was novel to the model. The trained model classified approximately 2,000 images per minute on a laptop computer with 16 gigabytes of RAM. The trained model achieved 98% accuracy at identifying species in the United States, the highest accuracy of such a model to date. Out-of-sample validation from Canada achieved 82% accuracy, and correctly identified 94% of images containing an animal in the dataset from Tanzania. We provide an R package (Machine Learning for Wildlife Image Classification; MLWIC) that allows the users to A) implement the trained model presented here and B) train their own model using classified images of wildlife from their studies. The use of machine learning to rapidly and accurately classify wildlife in camera trap images can facilitate non-invasive sampling designs in ecological studies by reducing the burden of manually analyzing images. We present an R package making these methods accessible to ecologists. We discuss the implications of this technology for ecology and considerations that should be addressed in future implementations of these methods.

The Mexican wolf (Canis lupus baileyi) was included in the 1973 Endangered Species Act listing of the gray wolf (C. lupus), but then listed separately as a subspecies in 2015. Early accounts of its range included the Sierra Madre Occidental of Mexico, southeastern Arizona, southwestern New Mexico, and sometimes western Texas, supported by ecological, biogeographic, and morphological data. There have been multiple unsuccessful attempts to revise the original 1982 recovery plan and identify areas suitable for Mexican wolf reintroduction. Despite the fact that 90% of its historical range is in Mexico and widespread suitable habitat exists there, previous draft recovery plans recommended recovery mostly outside of Mexico and well north of the subspecies' historical range. Planning recovery outside historical range of this subspecies is fraught with problems that may compromise, thwart, or impede successful recovery. Dispersal of Mexican wolves northward and continued movements southward by Northwestern wolves (C. l. occidentalis), along with allowing establishment of Mexican wolves north of their historical range before they are recovered, may lead to premature and detrimental intraspecific hybridization. Interbreeding of Northwestern wolves from Canadian sources and Mexican wolves does not represent the historical cline of body size and genetic diversity in the Southwest. If Northwestern wolves come to occupy Mexican wolf recovery areas, these physically larger wolves are likely to dominate smaller Mexican wolves and quickly occupy breeding positions, as will their hybrid offspring. Hybrid population(s) thus derived will not contribute towards recovery because they will significantly threaten integrity of the listed entity. Directing Mexican wolf recovery northward outside historical range threatens the genetic integrity and recovery of the subspecies, is inconsistent with the current 10(j) regulations under the ESA, is unnecessary because large tracts of suitable habitat exist within historical range, is inconsistent with the concepts of restoration ecology, and disregards unique characteristics for which the Mexican wolf remains listed.

There is substantial literature on the importance of bridging across disciplinary and science–management boundaries. One of the ways commonly suggested to cross boundaries is for participants from both sides of the boundary to jointly produce information (i.e., knowledge co-production). But simply providing tools or bringing people together in the same room is not sufficient. Here we present a case study documenting the mechanisms by which managers and scientists collaborated to incorporate climate change projections into Colorado’s State Wildlife Action Plan. A critical component of the project was the use of a collaborative modeling and visualization workspace: the U.S. Geological Survey’s Resource for Advanced Modeling (RAM). Using video analysis and pre/post surveys from this case study, we examine how the RAM facilitated cognitive and social processes that co-produced a more salient and credible end product. This case provides practical suggestions to scientists and practitioners who want to ...

1. Motion-activated cameras (“camera traps”) are increasingly used in ecological and management studies for remotely observing wildlife and have been regarded as among the most powerful tools for wildlife research. However, studies involving camera traps result in millions of images that need to be analyzed, typically by visually observing each image, in order to extract data that can be used in ecological analyses.2. We trained machine learning models using convolutional neural networks with the ResNet-18 architecture and 3,367,383 images to automatically classify wildlife species from camera trap images obtained from five states across the United States. We tested our model on an independent subset of images not seen during training from the United States and on an out-of-sample (or “out-of-distribution” in the machine learning literature) dataset of ungulate images from Canada. We also tested the ability of our model to distinguish empty images from those with animals in another out-of-sample dataset from Tanzania, containing a faunal community that was novel to the model.3. The trained model classified approximately 2,000 images per minute on a laptop computer with 16 gigabytes of RAM. The trained model achieved 98% accuracy at identifying species in the United States, the highest accuracy of such a model to date. Out-of-sample validation from Canada achieved 82% accuracy, and correctly identified 94% of images containing an animal in the dataset from Tanzania. We provide an R package (Machine Learning for Wildlife Image Classification; MLWIC) that allows the users to A) implement the trained model presented here and B) train their own model using classified images of wildlife from their studies.4. The use of machine learning to rapidly and accurately classify wildlife in camera trap images can facilitate non-invasive sampling designs in ecological studies by reducing the burden of manually analyzing images. We present an R package making these methods accessible to ecologists. We discuss the implications of this technology for ecology and considerations that should be addressed in future implementations of these methods.

Wildlife populations are often influenced by multiple political jurisdictions. This is particularly true for wide-ranging, low-density carnivores whose populations have often contracted and remain threatened, heightening the need for geographically coordinated priorities at the landscape scale. Yet even as modern policies facilitate species recoveries, gaps in knowledge of historical distributions, population capacities, and potential for genetic exchange inhibit development of population-level conservation priorities. Wolverines are an 8–18 kg terrestrial weasel (Mustelidae) that naturally exist at low densities (∼5/1000 km 2 ) in cold, often snow-covered areas. Wolverines were extirpated, or nearly so, from the contiguous United States by 1930. We used a resource selection function to (1) predict habitat suitable for survival, reproduction and dispersal of wolverines across the western US, (2) make a rough estimate of population capacity, and (3) develop conservation priorities at the metapopulation scale. Primary wolverine habitat (survival) existed in island-like fashion across the western US, and we estimated capacity to be 644 wolverines (95% CI = 506–1881). We estimated current population size to be approximately half of capacity. Areas we predicted suitable for male dispersal linked all patches, but some potential core areas appear to be relatively isolated for females. Reintroduction of wolverines to the Southern Rockies and Sierra-Nevadas has the potential to increase population size by >50% and these regions may be robust to climate change. The Central Linkage Region is an area of great importance for metapopulation function, thus warranting collaborative strategies for maintaining high survival rates, high reproductive rates, and dispersal capabilities. Our analysis can help identify dispersal corridors, release locations for reintroductions, and monitoring targets. The process we used can serve as an example for developing collaborative, landscape-scale, conservation priorities for data-sparse metapopulations.

Black-tailed prairie dogs (Cynomys ludovicianus) are a species of concern and accurate estimates of occupied area are required to assess their status. We conducted aerial line-intercept surveys to estimate colony areas in Colorado, USA, 2006–2007. Optimal allocation based on results from a previous (2002) survey was used to distribute flight time to sample 28 counties. Uncorrected estimates of active and inactive colony areas from the aerial surveys were 329,529 (SE = 16,841) ha and 18,292 (SE = 2,366) ha, respectively. We attempted to ground-truth a randomly selected sample of 186 colony intercepts but gained complete access to only 150. Ground-truthing demonstrated that aerial surveys estimated only 96% of the true lengths of colony intercepts but overestimated the proportion of active colonies. Corrected estimates of active and inactive colony areas are 319,165 (SE = 20,105) ha and 42,422 (SE = 11,485) ha, respectively. Because ground-truthing was not conducted in the original 2002 survey, unc...

Throughout the American West, private open lands are being converted to a human-dominated residential landscape. This conversion is the foremost threat to wildlife habitat. Pitkin County, Colorado, home to the City of Aspen, is an internationally known example of this land use change. In this article, we describe patterns of development in Pitkin County from the late 1800s through the 1990s. Using field data, we estimate the overall loss of songbird habitat due to two types of residential development, clustered and dispersed, and then examine the types of vegetation most affected. From these data we demonstrate that clustered development, in contrast to dispersed development, can reduce the impact of exurban residential development on wildlife habitat.

Residential development is occurring at unprecedented rates in the Rocky Mountain region of the United States, with unknown ecological consequences. We conducted our research in exurban development in Pitkin County, Colorado, between May and June in 1998 and 1999. Unlike suburban development, exur- ban development occurs beyond incorporated city limits, and the surrounding matrix remains the original ecosystem type. We surveyed songbirds and medium-sized mammals at 30, 180, and 330 m away from 40 homes into undeveloped land to examine the effect of houses along a distance gradient, and in developments of two different housing densities as well as undeveloped sites to examine the effect of housing density. We placed bird species into one of two groups for the house-distance effect: (1) human-adapted species, birds that occurred in higher densities close to developments and lower densities farther away and (2) human-sensitive species, birds that occurred in highest densities farthest from homes and in lowest densities close to develop- ment. For both groups, densities of individual species were statistically different between the 30- and 180-m sites. Six species were classified as human-adapted, and six were classified as human-sensitive for the house- distance effect. Dogs ( Canis familiaris ) and house cats ( Felis domesticus ) were detected more frequently closer to homes than farther away, and red foxes ( Vulpes vulpes ) and coyotes ( Canis latrans ) were detected more fre- quently farther away from houses. With respect to the effect of housing density, most avian densities did not differ significantly between high- and low-density development but were statistically different from undevel- oped sites. Six species were present in higher densities in developed areas, and eight species were present in higher densities in undeveloped parcels. Similar results were found for mammalian species, with dogs and cats detected more frequently in high-density developments and red foxes and coyotes detected more fre- quently in undeveloped parcels of land. From an ecological standpoint, it is preferable to cluster houses and leave the undeveloped areas in open space, as opposed to dispersing houses across the entire landscape.