Trail camera images, data, and scripts used to assess AI model classification of bighorn sheep and compare default model training to retrained versions

The files included herein are associated with the manuscript: “Improving AI Performance in Wildlife Monitoring through Species and Environment-Specific Training: A Case Study on Desert Bighorn Sheep” submitted to Ecological Informatics (Okuley et al. 2025). This project assessed existing AI model software designed to classify wildlife species in images for their ability to correctly identify desert bighorn sheep. Collaborators from the National Park Service (NPS) and California Department of Fish and Wildlife (CDFW) previously conducted extensive camera surveys of water sources throughout desert regions in California to detect bighorn sheep presence and abundance. Manually classifying these images, however, presents a bottleneck and challenge to efficiently use survey data to make decisions. We sampled a training and testing set of trail camera images from NPS/CDFW surveys that occurred over 2016 - 2023, classified images as having bighorn sheep or no bighorn sheep, and ran two AI model classifiers on testing images. Based on performance metrics and a generalized linear mixed model (glmm) analysis, we retrained the top performing classifier with a new set of targeted training images and tested for performance gains. The R code and .csv data can be used to recreate figures describing results, calculate all AI classifier performance metrics presented in the paper, and carry out the generalized linear mixed model (glmm) analysis that assessed performance between classifiers and across different metrics of image quality. All images used as the testing dataset (n = 95,547) and to retrain (n = 13,000) the AI models are also included and have been reviewed and classified as containing "bighorn sheep" or "no bighorn sheep" by a human researcher.