Publication Date

Spring 2021

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Environmental Studies

Advisor

Lynne Trulio

Keywords

camera trap, convolutional neural network, environmental technology, invasive species, machine learning applications, wildlife conservation

Subject Areas

Environmental studies; Computer science; Wildlife conservation

Abstract

Convolution neural network models (CNNs) can successfully identify animal species in camera-trap images in simplified testing environments. CNN performance in more complex, realistic environments is understudied. Here the Wellington Camera Traps dataset was used to simulate a wildlife conservation project to detect invasive species at low population levels using camera-trap images and CNN models. Ten CNNs were developed and analyzed with seven testing datasets, simulating 13 possible project scenarios. Model performance was measured using standard computer science metrics, top-1, and top-5 accuracy, and two novel performance metrics developed for this research to directly reflect wildlife conservation goals, false alarm rate, and missed invasive rate. The highest performing models achieved 91.8% and 99.6% top-1 and top-5 accuracy; however, these models also had the highest missed invasive rates. This effect was related to the ratio of native to invasive species in the model’s training images. As this ratio increased so did the model’s top-1 and top-5 accuracy but also the missed invasive rate. Thus to achieve optimal performance when selecting or training a CNN for use in a wildlife camera-trap project the metric used to judge the performance of the model must be tailored to the specific goals of the project, and the distribution of species in the model’s training images must match the distribution that will be seen in the project’s camera-trap images.

Download

Share

COinS