An analysis of altitude, citizen science and a convolutional neural network feedback loop on object detection in Unmanned Aerial Systems

摘要

Using automated processes to detect wildlife in uncontrolled outdoor imagery in the field of wildlife ecology is a challenging task. In imagery provided by Unmanned Aerial Systems (UAS), this is especially true where individuals are small and visually similar to background substrates. To address these challenges, this work presents an automated feedback loop which can operate on large scale imagery, such as UAS generated orthomosaics, to train convolutional neural networks (CNNs) with extremely unbalanced class sizes. This feedback loop was used to help train CNNs using imagery classified by both expert biologists and citizen scientists at the Wildlife@home project. Utilizing the feedback loop dramatically reduced population count error rates from previously published work: from +150% to 3.93% on citizen scientist training data and +88% to +5.24% on expert training data. The system developed was then utilized to investigate the effect of altitude on CNN predictions. The training dataset was split into three subsets depending on the altitude of the imagery (75 m,100 m and 120 m). While the lowest altitude was shown to provide the best predictions of the three (+11.46%), the aggregate data set still provided the best results (-3.93%) indicating that there is greater benefit to be gained from a large data set at this scale, and there is potential benefit to having training data from multiple altitudes. This article is an extended version of "Detecting Wildlife in Unmanned Aerial Systems Imagery using Convolutional Neural Networks Trained with an Automated Feedback Loop" published in the proceedings of the 18th International Conference of Computational Science (ICCS 2018) [1]. (C) 2019 Elsevier B.V. All rights reserved.