Machine learning for camera trap image classification

Supervisor: Dr Peter Stewart and Dr Tiffany Vlaar

School: Mathematics & Statistics

Description: 

Motion-activated cameras – known as camera traps or trail cameras – enable the monitoring of rare and elusive animals in remote locations over all hours of the day. As a result, they have become an indispensable tool in applied ecology and conservation. However, a key challenge is that camera traps can produce thousands of images per day; identifying the species in these images can take many hours, and may require expert knowledge if species are difficult to distinguish or the images are poor quality. In recent years, convolutional neural networks (CNNs) have been deployed to aid in this classification process. However, it is not yet clear how the performance of CNNs compares to that of human classifiers, and how this performance varies according to factors such as the type of species and the quality of the image.  

Prickly Pear Project Kenya (https://www.zooniverse.org/projects/peter-dot-stewart/prickly-pear-project-kenya) used camera traps to study the impact of invasive prickly pear cacti (Opuntia spp.) on wild mammals in Laikipia County, Kenya. The dataset comprises images of more than 40 species, ranging in size from dik-dik (one of the smallest species of antelope) to the African elephant. During the project, around 187,000 camera trap images were classified by more than 8000 volunteers; each image was classified by multiple volunteers, and a subset of around 27,000 images were also classified by an expert. Consequently, the Prickly Pear Project Kenya dataset provides an excellent opportunity to compare the accuracy of classifications made by a CNN to those made by human classifiers.  

 The aims of this project are to: 

  1. Train CNNs to classify species in the Prickly Pear Project Kenya camera trap dataset 
  2. Compare the accuracy of classifications made by the CNNs to those made by human classifiers 
  3. Evaluate CNN performance on images that are difficult to classify for humans (i.e., those with substantial disagreement between human classifiers) 
  4. If time allows, explore how CNN performance varies under different neural network optimisation strategies