|Aerial Wildlife Image Repository (AWIR) – a Dataset to Accelerate the Development of Highly Accurate Machine Learning Models to Classify Wildlife from Aerial Imagery|
|Sathishkumar Samiappan, Jared Elmore, Meilun Zhou, Morgan Pfeiffer, Bradley Blackwell, Raymond Iglay, Kristine Evans|
The increased use of small (<55lbs) uncrewed aircraft systems (UAS) to survey wildlife presents a need and opportunity for researchers to train computer vision algorithms on UAS imagery regarding animal detection, especially wildlife species. Computer vision and machine learning models trained on aerial wildlife imagery will enable enhanced ability to detect and accurately count wildlife. This capability is important among wildlife conservation and management as well as human-wildlife conflict resolution. However, the availability of aerial wildlife imagery for training machine learning models is relatively non-existent. As such, we propose an Aerial Wildlife Image Repository (AWIR), a suite of opensource datasets which will be annotated from both aerial images and videos (hereafter referred to as data) captured from UAS. These data will be available to train machine learning and computer vision algorithms that can detect wildlife from aerial imagery. Currently, no repository exists publicly for this purpose, and AWIR would be the first of its kind. Computer vision algorithms for recognizing objects in images learn through large amounts of data (supervised learning). Therefore, a primary purpose of AWIR will be improving computer recognition of wildlife in images captured from UAS. This poster will introduce the audience to the initial version of AWIR, the details of steps involved in creating this dataset, results from state-of-the-art computer vision algorithms applied on AWIR and future plans.
|The Unpublished Realities of Wildlife Monitoring with Unoccupied Aerial Systems (UAS): A Survey of End Users|
|Raymond Iglay, Jared Elmore, Meilun Zhou, Kristine Evans, Sathishkumar Samiappan, Morgan Pfeiffer, Bradley Blackwell|
Small (< 55 lbs) Unoccupied Aerial Systems (UAS) have demonstrated promise as wildlife monitoring tools for conservation, management, and research endeavors. Despite their use over the past 20 years, only recently have coordinated research efforts emerged evaluating UAS efficiencies and associated standard operating procedures. Therefore, most information regarding UAS wins, fails, and potential is still institutional knowledge by field practitioners. Complementing efforts to improve UAS wildlife monitoring from field sampling through computer vision applications, we surveyed over fifty wildlife professionals with interest in UAS applications such as members of The Wildlife Society’s Drone Working Group and research collaborators in academia and government to gain a better understanding of how UAS have been or are being used in field applications. Over 80% of respondents were from academia or a federal government agency with just over half (57%) having pilot and data analysis responsibilities and at least one peer-reviewed publication regarding UAS. Near equal interest in monitoring birds (42%) and mammals (35%) also emerged, but analysis of bird imagery well surpassed mammals (53% vs. 35%). Additional summaries regarding UAS sensor types and performance, UAS platforms, time dedicated to UAS missions, animal sizes of target species, and UAS sampling issues were also reviewed and provide more insights to UAS wildlife monitoring considerations for current and future field endeavors. Rapid technology advancements and an inherent lag in technology adoption are not new, nor unique to UAS and the wildlife profession. However, information gained from this survey demonstrates the limits of peer-reviewed information sources when developing future research directives and the value of stakeholder cooperation.
|Precision of Demographic Ratios from Aerial Surveys of Large Mammals: Number of Detections Matters|
|Randy W. DeYoung, Aaron Foley, Landon Schofield, David G. Hewitt, Tyler Campbell|
Aerial surveys are often used to collect data for management purposes. Ratios are crucial components of surveys and are used to determine composition relative to management objective. However, there is little guidance on minimum number of detections needed to obtain precise demographic ratios from aerial surveys. We evaluated precision of demographic ratios during aerial surveys of white-tailed deer (Odocoileus virginianus) on rangelands in South Texas, USA. We used 4-seat helicopters to conduct fixed-width transect surveys during 2011–2015 and 2018–2020. We used the methods of Czaplewski et al. (1983) to calculate precision of fawn:doe ratios (fawns/100 does) and buck:doe ratios (bucks/100 does) at 3 spatial scales: pasture (n = 617, x̄ = 2,440 ha), wildlife management unit (WMU, n = 188, x̄ = 8,120 ha), and ranch (n = 15, x̄ = 17,650 ha). A ratio was considered to be precise when the confidence interval was ≤10. Only 18% and 6% of pasture-year surveys generated precise fawn:doe and buck:doe ratios, respectively. At the WMU-year scale, frequency of precise estimates increased for fawn:doe (46%) and buck:doe (24%) ratios. At the ranch scale, all fawn:doe ratios and 93% of buck:doe ratios were precise. Our results indicate that a large spatial scale is needed to obtain precise ratios via helicopter surveys in South Texas which is likely a function of relatively low deer densities (~10 ha/deer) in conjunction with detection probabilities of ~0.50. However, on average, both ratios approximated ratios at the end of the survey when ≥20–30 adult does were observed. Therefore, trends over time may be more valuable than current-year numbers for relatively small spatial scales (<8,000 ha). Our results also reinforce the importance of quantifying precision to determine if changes in population demographics were significant or not and whether survey effort needs to be adjusted.
|Should Distance Sampling Methods Be Used When Surveying Wildlife Via Drones?|
|Jesse Exum, Aaron Foley, David G. Hewitt, Randy W. DeYoung, Jeremy Baumgardt, Mickey Hellickson|
Drones have emerged as another tool in the toolbox of wildlife research and management. One application of drones is to survey wildlife but the limited number of previous drone-based population estimates assume no visibility bias; thus, uncorrected (raw) counts are often reported. We tested the hypothesis that the uniform function in distance sampling analysis would not best fit the distribution of distances which would indicate that probability of detection declines at a certain distance. In South Texas, we used quadcopters equipped with thermal video cameras to conduct repeated daytime surveys for white-tailed deer (Odocoileus virginianus) on 3 study sites during February–April 2020. Drones were programmed to fly fixed-width transects 37 m above ground level at 24 km/hr which generated a 57 m swath. Results from sites 1 and 2 best fit the hazard rate key function with an average probability of detection of 0.63 (range = 0.56–0.70). Probability of detection appreciably declined at ~20-25 m from the transect. Site 3 had a negative-skewed distance distribution because the 80% herbaceous cover (vs 42–43% herbaceous cover in sites 1 and 2) caused severe solar reflectance in the center of the footage. Coefficient of variations of uncorrected and corrected estimates were similar in site 1 (n = 5, CV = 9% vs 12%, respectively) but variation in corrected estimates was lower in site 2 (n = 11, CV = 41% vs 20%). Overall, our results indicate that probability of detection was not uniform during daytime thermal surveys; therefore, uncorrected counts should be avoided. Further, the use of distance sampling may result in more consistent estimates from repeated surveys but application may be limited to certain habitat types.
|Using Remote Sensing to Quantify Habitat for a Declining Grassland Bird Species – SRIP|
|Sierra Moore, Elizabeth Hunter, Abbie Dwire|
|Henslow’s Sparrows are a grassland bird species of conservation concern that winter in the Southeastern United States. Henslow’s Sparrow populations have been declining over recent decades, along with many other grassland birds. Henslow’s Sparrows wintering in Southeast Georgia select microhabitats with specific herbaceous plant structure. The objective of this study is to connect known Henslow’s Sparrow microhabitat features in Southeast Georgia to remotely-sensed vegetation features to identify suitable Henslow’s Sparrow habitat at a broader scale. To do this, we plan to analyze satellite image texture to identify habitat features, such as vegetation structure and compositional heterogeneity, and connect those habitat features to microhabitats that Henslow’s Sparrows use. Previous research has estimated Henslow’s Sparrow microhabitat use by comparing vegetation plots in used areas (estimated through telemetry) to unused areas. We will use ERDAS Imagine software and Google Earth Engine to analyze remotely sensed Normalized Difference and Enhanced Vegetation Indices (NDVI, EVI) to calculate image texture metrics for the same vegetation plots to determine whether these remotely sensed metrics correlate with important structural microhabitat features of plant height and density. We will then calculate relevant texture metrics for the species range in Georgia to identify potentially suitable habitat areas across the state. We expect our results to align with previous research demonstrating that satellite image texture can be used to estimate grassland bird habitat occurrence. This research will allow managers to better identify suitable habitat for conservation and management across the species’ wintering range.|
|Evaluating the Use of Drones for Locating Grassland Nesting Ducks – SRIP|
|Cailey Isaacson, Kaylan Kemink, Kyle Kuechle, Catrina Terry, Javier Lenzi, Susan Ellis-Felege|
|Drones have become a popular tool among wildlife biologists as they have a variety of applications and have been proposed for wildlife surveys. Traditionally, duck nests are commonly located and monitored using a very labor-intensive approach of ATV chain dragging (Klett et al. 1986). The ATV chain dragging approach can be invasive to the landscape leaving ATV tracks and often disturbing other grassland nesting species and native vegetation. Thermal sensors fitted on drones has been proposed as a possible tool for locating waterfowl nests, but it can be intensive and challenging (Bushaw et al. 2020, Helvey et al. 2020). One potential challenge is that thermal sensors provide relative temperatures gradients that require a contrast to differentiate between the target of interest and the background. Thus, there is a need to determine the effectiveness of these for locating nests in terms of search time, detection, and optimal protocols. The objectives of this research in progress includes 1) Do thermal sensors on drones detect duck nests at the same rate as traditional ATV searches, and 2) Is nest success higher on plots searched using drones than those searched using ATVs? Field methods include locating and monitoring nests on three plots that are searched by both ATV and drone, three plots using only an ATV, and three plots using only the drone. On plots that are searched by both methods, flights will occur within 24 hours of an ATV search. Warmer locations found on thermal imagery will be checked to confirm as a true positive or false positive nest detection. Weather information, and nesting data will be recorded for all monitored nests and flights. All nests will be monitored under similar conditions until nest hatch or fail. Resulting data will calculate detection rates between survey methods and probability of nest failure.|
|Using Unmanned Aircraft Systems (UAS) to Assess Crop Damage by Wild Pigs in Alabama – SRIP|
|Arielle Fay, Mark Smith, Wesley Anderson, Stephen Ditchkoff|
|Wild pigs (Sus scrofa) are a non-native species causing significant damage annually to agriculture in the United States. Despite previous and ongoing research documenting wild pig damage to agriculture and natural resources, few studies in the United States have examined the spatial and temporal changes in the amount and distribution of wild pig damage occurring within row crop fields throughout the growing season. Additionally, no studies examined how surrounding landscape characteristics may predispose some fields to damage. Therefore, we used an unmanned aircraft system (UAS; hereafter drone) coupled with a multispectral sensor to quantify the amount, distribution, and timing of wild pig damage to corn, soybean, and peanut fields on private agriculture production farms in southwest and southeast Alabama during 2021 (and again in 2022). We conducted drone flights over approximately 27 row crop fields about every 2-3 weeks from planting to harvest and then conducted infield inspections the following day to verify the damage was caused by wild pigs while using a handheld GPS unit to map the spatial extent of the damage. We then conducted image classifications based on spectral reflectance to delineate damage vs undamaged portions of the field and compared these estimates to those obtained through infield measurements. We then used FRAGSTATS to compute patch, class, and landscape-level metrics of land cover types with an 800-meter buffer of each field to develop predictive models. We present the first year (2021) results of this study. The results of this study will be used to better understand the temporal and spatial depredation patterns of wild pigs in agriculture landscapes and to guide the development of management actions targeted to reduce damage.|
|A Remote Sensing and Radio Telemetry Biosecurity Mechanism for the Poultry Industry – SRIP|
|Matt Hardy, Chris Williams, Jeff Buler, Brian Ladman, Michael Casazza, Maurice Pitesky, Cory Overton, Elliott Matchett|
|The risk of avian influenza virus to commercial poultry operations may be increased due to interactions with wild waterfowl. We are examining the potential use of combined data streams to create real-time mapping products which mitigate that risk. First, using high frequency GPS/GSM-telemetry and Brownian bridge movement models, we will generate wintering distribution maps for target waterfowl species (Mallard, Canada Goose, and Greater Snow Goose) in the mid-Atlantic and California. Further, we are determining if temporal and spatial differences exist in movement patterns as a function of environmental variables, resource depletion, and anthropogenic pressure (e.g., hunting seasons) via data collected from GPS/GSM-telemetry marked waterfowl. Second, we are validating a novel approach to quantify waterfowl density in the airspace within the Delmarva Peninsula (Delaware, Maryland, and Virginia), North Carolina, and California during November and March, 2020–2022 to pinpoint areas of high potential AIV contact at poultry farms in relation to overall waterfowl density in the airspace using data from the NEXRAD weather surveillance radar network. Weather surveillance radar techniques provide a more comprehensive assessment of bird activity both in the airspace, and on the ground, at the onset of feeding flight exodus. Preliminary research has shown that the patterns observed from both weather radar and GPS/GSM-telemetry techniques are closely correlated. Combining these data streams, we ultimately aim to create a near real time interactive map that will be easily accessible to poultry farmers and act as a biosecurity mechanism for the poultry industry.|