Improving AI performance in wildlife monitoring through species and environment-specific training: A case study on desert Bighorn sheep

Abstract

Motion-activated cameras are widely used to monitor wildlife, offering a non-intrusive and cost-effective means to collect high volumes of data. Artificial intelligence (AI) models can expedite image processing, but automated species classifications can be too inaccurate to meet end-users' needs. This study evaluates how selection of data for model training influences AI detection of a focal species (desert bighorn sheep; Ovis canadensis nelsoni) across similar and novel locations. We compared two AI models: a species-specialist (deep_sheep) and a species-generalist (CameraTrapDetectoR), identified sources of bias, and retrained the specialist model using two datasets targeted toward biases associated with classification failure. Testing on 95,547 images from 36 water sources (5 novel) in the Mojave and Sonoran Deserts revealed the specialist model outperformed the generalist by 21.44 % in accuracy and reduced false negatives by 45.18 %. The specialist model was retrained first on site-representative data, then on both site-representative and extreme image-condition data. Retraining iterations consecutively reduced the false negative rate (36.94 % → 6.23 % → 4.67 %) and improved reliability across sites at the cost of a reciprocal increase in false positive rate (2.87 % → 15.22 % → 23.97 %) and variation. The site-representative retraining had the highest overall accuracy. Accuracy at out-of-sample sites remained comparable to the full dataset, though minor performance declines were observed. We found that specifying an AI's training to single-species classification and conditions within specific environments produced robust classification accuracy at minimal data requirements. By narrowing objectives while ensuring adequate training data variety, we achieved 89.33 % accuracy with a small fraction of the training data required by similar performing models.