Program on the Environment

Monitoring tagged and branded animals is vital for the conservation of a species, and one of the most popular ways to monitor these individuals is with camera traps. Camera traps made to monitor these animals produce thousands of images, which is time-consuming for researchers and conservationists; machine learning algorithms can go through this data in an efficient manner, saving researchers’ time. The aim of this study was to identify which environmental factors most contributed to errors in the machine learning model, aimed at identifying branded Steller sea lions. During my time interning with NOAA training this model, I compiled a data set in which I noted images in which the model made a mistake and created a dataset highlighting causes, sex, age, and what the image was flagged as. Utilizing my dataset, I found that out of the three months I had data for, September, which is closest to breeding season, saw the most errors. Of the confounding factors, environmental aspects and camera obstruction were the causes of most errors. By identifying confounding factors in a machine learning model, researchers can improve generalization, reduce bias, and create better training datasets. Shortening training time for the machine learning model further gives researchers time to focus on the conservation and management of the western Steller sea lion population.