Application of unoccupied aerial vehicle (UAV) thermal imagery and use of deep learning to survey wild turkey populations
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
Although wild turkeys (Meleagris gallopavo) are an important game species, it is difficult to estimate abundance at spatial scales relevant for management. Unoccupied aerial vehicles (UAV) may provide a means to efficiently survey turkeys on large spatial scales. The objectives of this study were to determine if roosting wild turkeys can be detected in thermal video collected from UAVs by observers and an automated deep learning model and what habitat and environmental conditions influence automated detection probability. I evaluated if observers and deep learning can differentiate between turkeys and vultures. Finally, I determined survey effort required to detect a 5-20 percent change in wild turkey population size over 5 years in landscapes characteristic of Texas Wildlife Management Districts 3, 5, and 6. I collected thermal video from UAV flights over roost locations of GPS-tagged wild turkeys (n =30 roosts) and opportunistically over 6 turkey vulture (Cathartes aura) roosts in three Management Districts of Texas in winter 2021 and 2022. I applied a trained single-class deep learning model (U-Net) to UAV thermal videos of roosting wild turkeys (n = 62). Using logistic regression, I evaluated the effect of habitat and environmental conditions on deep learning detection probability and false positive rate. I created a survey and asked fifteen practicing biologists to count all roosting birds and identify their species (turkeys or vultures) in 36 video clips (1-2 min) that contained 0 - 67 birds. To determine if deep learning could differentiate between turkeys and vultures, I applied a trained, multi-class deep learning model (U-Net) to UAV thermal videos of roosting wild turkeys (n = 26) and vultures (n = 6). I used a series of simulations to determine survey effort required to detect a 5-20 percent change in turkey populations over five years based on detections from the deep learning model. Overall, detection probability for the trained U-Net model was 0.71 (SD = 0.34) which was influenced by temperature and humidity at the time of flight. Mean false positive rate was 0.53 (SD = 0.34) and was influenced by temperature at the time of flight and canopy height. Biologists tended to overcount by an average of 9 percent (SD = 1.47) when video clips included [less than] 35 turkeys and undercount by an average of 17 percent (SD = 0.22) when video clips included 35 turkeys. Differentiating between turkeys and vultures proved difficult for both biologists and the trained U-Net model. Such biases may result in inaccurate population estimates if not accounted for. Simulations demonstrated that the trained U-Net model applied to thermal UAV video would underestimate population abundance by 36.13 percent (SD = 19.26) on average. Future work should focus on improving deep learning networks to increase detection probability and decrease false positive detections. I found that there was insufficient power ([greater than] 0.80) to detect a 5 percent change in turkey populations over 5 years using thermal UAV surveys and the deep learning model created for this study but there is enough power to detect population changes [greater than or equal to] 10 percent. To maximize success of this method weather conditions should be carefully considered prior to flights and managers should aim to complete surveys in temperature [less than] 10oC and [less than] 50 percent humidity. UAVs equipped with thermal cameras offer a promising novel method to survey roosting wild turkeys and estimate population sizes which will contribute to more effective management of the species.
Table of Contents
PubMed ID
Degree
M.S.