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/.

Fusion of visible and thermal images improves automated detection and classification of animals for drone surveys.

Visible and thermal images acquired from drones (unoccupied aircraft systems) have substantially improved animal monitoring. Combining complementary information from both image types provides a powerful approach for automating detection and classification of multiple animal species to augment drone surveys. We compared eight image fusion methods using thermal and visible drone images combined with two supervised deep learning models, to evaluate the detection and classification of white-tailed deer (Odocoileus virginianus), domestic cow (Bos taurus), and domestic horse (Equus caballus). We classified visible and thermal images separately and compared them with the results of image fusion. Fused images provided minimal improvement for cows and horses compared to visible images alone, likely because the size, shape, and color of these species made them conspicuous against the background. For white-tailed deer, which were typically cryptic against their backgrounds and often in shadows in visible images, the added information from thermal images improved detection and classification in fusion methods from 15 to 85%. Our results suggest that image fusion is ideal for surveying animals inconspicuous from their backgrounds, and our approach uses few image pairs to train compared to typical machine-learning methods. We discuss computational and field considerations to improve drone surveys using our fusion approach.

Considering the influence of land use/land cover on estuarine biotic richness with Bayesian hierarchical models.

The composition of land use/land cover (LULC) in coastal watersheds has many implications for estuarine system ecological function. Land use/land cover can influence allochthonous inputs and can enhance or degrade the physical characteristics of estuaries, which in turn affects estuaries' ability to support local biota. However, these implications for estuaries are often poorly considered when assessing the value of lands for conservation. The focus of research regarding terrestrial and estuarine interfaces often evaluates how LULC may stress estuarine ecosystems, but in this study we sought to understand how LULC may both positively and negatively affect estuaries using measures of observed biotic richness as proxies for estuarine function. We investigated the influence of LULC on estuarine biotic richness with Bayesian hierarchical models using multiple geospatial data sets from 33 estuaries and their associated watersheds along the Gulf of Mexico coastal region of the United States. We designed the hierarchical models with observed species richness of three functional groups (FGs) (i.e., pelagic fishes, forage fishes, and shrimp) from fishery-independent trawl surveys as response variables. We then set salinity and water temperature as trawl-specific covariates and measures of influence from six LULC classes as estuary-specific covariates and allowed the models to vary by estuary, trawl program, salinity, and temperature. The model results indicated that the observed richness of each FG was both positively and negatively associated with different LULC classes, with estuarine wetlands and forested lands demonstrating the strongest positive influences on each FG. The results are generally consistent with past studies, and the modeling framework provides a promising way to systematically quantify LULC linkages with the biotic health of estuaries for the purposes of potentially valuing the estuarine implications of land conservation.

Improving Animal Monitoring Using Small Unmanned Aircraft Systems (sUAS) and Deep Learning Networks.

In recent years, small unmanned aircraft systems (sUAS) have been used widely to monitor animals because of their customizability, ease of operating, ability to access difficult to navigate places, and potential to minimize disturbance to animals. Automatic identification and classification of animals through images acquired using a sUAS may solve critical problems such as monitoring large areas with high vehicle traffic for animals to prevent collisions, such as animal-aircraft collisions on airports. In this research we demonstrate automated identification of four animal species using deep learning animal classification models trained on sUAS collected images. We used a sUAS mounted with visible spectrum cameras to capture 1288 images of four different animal species: cattle (Bos taurus), horses (Equus caballus), Canada Geese (Branta canadensis), and white-tailed deer (Odocoileus virginianus). We chose these animals because they were readily accessible and white-tailed deer and Canada Geese are considered aviation hazards, as well as being easily identifiable within aerial imagery. A four-class classification problem involving these species was developed from the acquired data using deep learning neural networks. We studied the performance of two deep neural network models, convolutional neural networks (CNN) and deep residual networks (ResNet). Results indicate that the ResNet model with 18 layers, ResNet 18, may be an effective algorithm at classifying between animals while using a relatively small number of training samples. The best ResNet architecture produced a 99.18% overall accuracy (OA) in animal identification and a Kappa statistic of 0.98. The highest OA and Kappa produced by CNN were 84.55% and 0.79 respectively. These findings suggest that ResNet is effective at distinguishing among the four species tested and shows promise for classifying larger datasets of more diverse animals.

Relationships between survival and habitat suitability of semi-aquatic mammals.

Spatial distribution and habitat selection are integral to the study of animal ecology. Habitat selection may optimize the fitness of individuals. Hutchinsonian niche theory posits the fundamental niche of species would support the persistence or growth of populations. Although niche-based species distribution models (SDMs) and habitat suitability models (HSMs) such as maximum entropy (Maxent) have demonstrated fair to excellent predictive power, few studies have linked the prediction of HSMs to demographic rates. We aimed to test the prediction of Hutchinsonian niche theory that habitat suitability (i.e., likelihood of occurrence) would be positively related to survival of American beaver (Castor canadensis), a North American semi-aquatic, herbivorous, habitat generalist. We also tested the prediction of ideal free distribution that animal fitness, or its surrogate, is independent of habitat suitability at the equilibrium. We estimated beaver monthly survival probability using the Barker model and radio telemetry data collected in northern Alabama, United States from January 2011 to April 2012. A habitat suitability map was generated with Maxent for the entire study site using landscape variables derived from the 2011 National Land Cover Database (30-m resolution). We found an inverse relationship between habitat suitability index and beaver survival, contradicting the predictions of niche theory and ideal free distribution. Furthermore, four landscape variables selected by American beaver did not predict survival. The beaver population on our study site has been established for 20 or more years and, subsequently, may be approaching or have reached the carrying capacity. Maxent-predicted increases in habitat use and subsequent intraspecific competition may have reduced beaver survival. Habitat suitability-fitness relationships may be complex and, in part, contingent upon local animal abundance. Future studies of mechanistic SDMs incorporating local abundance and demographic rates are needed.

Assessing multiregion avian benefits from strategically targeted agricultural buffers.

Mounting evidence of wildlife population gains from targeted conservation practices has prompted the need to develop and evaluate practices that are integrated into production agriculture systems and targeted toward specific habitat objectives. However, effectiveness of targeted conservation actions across broader landscapes is poorly understood. We evaluated multiregion, multispecies avian densities on row-crop fields with native grass field margins (i.e., buffers) as part of the first U.S. agricultural conservation practice designed to support habitat and population recovery objectives of a national wildlife conservation initiative. We coordinated breeding season point transect surveys for 6 grassland bird species on 1151 row-crop fields with and without native grass buffers (9-37 m) in 14 U.S. states (10 ecoregions) from 2006 to 2011. In most regions, breeding season densities of 5 of 6 targeted bird species were greater in the 500-m surrounding survey points centered on fields with native grass buffers than in landscapes without buffers. Relative effect sizes were greatest for Northern Bobwhite (Colinus virginianus), Dickcissel (Spiza americana), and Field Sparrow (Spizella pusilla) in the Mississippi Alluvial Valley and Eastern Tallgrass Prairie regions. Other species (e.g., Eastern Meadowlark [Sturnella magna], Grasshopper Sparrow [Ammodramus savannarum]) exhibited inconsistent relative effect sizes. Bird densities on fields with and without buffers were greatest in the Central Mixed-grass Prairie region. Our results suggest that strategic use of conservation buffers in regions with the greatest potential for relative density increases in target species will elicit greater range-wide population response than diffuse, uninformed, and broadly distributed implementation of buffers. We recommend integrating multiple conservation practices in broader agricultural landscapes to maximize conservation effectiveness for a larger suite of species.