Needle in a haystack: towards the use of machine learning to detect Texas tortoises in large datasets

Abstract

Texas tortoises (Gopherus berlandieri) are an elusive reptile native to south-central Texas and northern Mexico. Unlike their congenerics in North America, the Texas tortoise does not create a conspicuous burrow that can be monitored. Instead, they create a shallow depression (known as a pallet) often in thick vegetation that obscures the tortoise from a surveyor’s view. Camera traps provide an alternative to visual encounter surveys that allows for a greater number of survey hours without disturbing the environment. Because Texas tortoises are ectothermic, time-lapse mode must be used, which generates massive datasets. Machine learning provides an opportunity to streamline the identification of tortoises in these datasets. Our goal was to train machine learning models in the identification of Texas tortoises across key representatives of their habitat types and make the model available for use. We have established grids of camera traps across the range of the Texas tortoise. In Python, we developed a deep-learning model for image segmentation. Given an input image, the trained model detects tortoises and draws a polygon. We used CVAT.ai to annotate the images by drawing bounding polygons around tortoises in each image. The annotated training images were used for transfer learning with a pre-existing computer vision model which can enable skillful detecting without requiring a massive set of annotated training examples. We used precision-recall curves to evaluate performance. We then tested the models using a test dataset that had been excluded from the training process. The top performing model had a precision score of 0.93, recall of 0.83, and accuracy of 0.86 on the test dataset. This machine learning model, which has been made available online, could make trail cameras a viable option to detect Texas tortoises. This model provides a valuable resource to streamline post trail camera data-collection which can allow for greater monitoring efforts of the Texas tortoise and support conservation efforts.