The following schedule is from the 72nd Annual Conference of the Southeastern Association of Fish & Wildlife Agencies which was held October 21-24, 2018 in Mobile, Alabama.
AUTHORS: Briana D. Stewart, Auburn University, Alabama Cooperative Research Unit; James B. Grand, U.S. Geological Survey, Alabama Cooperative Research Unit; Carolyn Moore, Auburn University, Alabama Cooperative Research Unit
ABSTRACT: Estimating eastern wild turkey (Meleagris gallopavo silvestris; hereafter turkeys) population demographics precisely and accurately is essential for making effective harvest and habitat management plans. Demographic estimates once based on expert opinion or harvest data are now being collected through game camera surveys that can be repeated across space and time. However, game camera surveys usually result in large numbers of images that must be interpreted in a timely manner. Classifying these images based on expert review can be time-consuming, costly, and error-prone. To address these issues, we developed a model using supervised classification and machine learning in MATLAB (Mathworks, Inc.) to determine the presence of turkeys in images. The models were trained using 500-point features from 3,342 training images that were collected on two study areas at 44 locations and manually interpreted. We compared 23 image classification methods; the top five methods were: cubic support vector machine (SVM, 90.9% accuracy), quadratic SVM (90.3% accuracy), ensemble subspace k-nearest neighbor (KNN, 90.2% accuracy), fine KNN (89.9% accuracy), and medium gaussian SVM (88.0% accuracy). Cubic SVM was the most accurate method with omission rate of 4.04% and commission rate of 5.09%. The use of machine learning will greatly reduce the time required to interpret the thousands of photos that are often collected in game camera surveys, and with appropriate training data could be extended to other species of wildlife.
