loading page

The successes and pitfalls: Deep learning effectiveness in a Chernobyl field camera trap application
  • Rachel Maile,
  • Matthew Duggan,
  • Timothy Mousseau
Rachel Maile
University of South Carolina
Author Profile
Matthew Duggan
Cornell University
Author Profile
Timothy Mousseau
University of South Carolina
Author Profile

Abstract

Camera traps have recently become in-situ sensors for collecting information on animal abundance and occupancy estimates. When deployed over a large landscape, camera traps have become ideal for measuring the health of ecosystems, particularly in unstable habitats where it can be dangerous or even impossible to observe with conventional methods. However, manual processing of imagery is extremely time and labor intensive. Due to this, many studies have started to employ the use of machine learning tools, such as convolutional neural networks (CNNs), with the presumption that a large number of images (millions) is needed to devise an effective identification or classification model. We examined specific factors pertinent to camera trap placement in the field that may influence the accuracy metrics of a deep learning model that has been trained with a small set of images. False negatives and false positives may occur due to a variety of reasons that make it hard even for a human observer to classify, including local weather patterns and degree of light present. We trained a CNN to detect 16 different object classes (14 animal species, humans, and fires) across 9,576 images taken from camera traps placed in the Chernobyl Exclusion Zone. After analyzing wind speed, cloud cover, and temperature, there was a significant relationship with CNN error and temperature. Furthermore, we found that the model was more successful when images were taken during the day as well as when precipitation is not present. Given the relationship between the influencing variables studied and model error rates, it will be important to obtain a wider breadth of events in differing weather and daytime factors for future classification models. This study suggests further exploration into the causes of error in classification modeling is necessary given the unique challenges posed by the analysis of camera trap imagery.
01 Sep 2022Submitted to Ecology and Evolution
01 Sep 2022Assigned to Editor
01 Sep 2022Submission Checks Completed
06 Sep 2022Reviewer(s) Assigned