Hyperparameter Tuning Improves Computer Vision Tools for Wildlife Conservation

Abstract

Camera traps are an excellent way to collect a large amount and variety of wildlife data, however, the volume of images poses a serious challenge for manual analysis. Through the use of machine learning and computer vision, this process can be automated, enabling rapid and accurate classification and identification, thus aiding conservation efforts. This study explores the effectiveness of EfficientNet models in classifying eight different classes from camera trap images, focusing on how hyperparameters such as learning rate and batch size affect model performance. Using a dataset of 16,487 images from Tai National Park, we experimented with different hyperparameter values and found that a lower learning rate and a moderate batch size yielded the highest accuracy. The efficientnet_b1 model was the most effective model, achieving 86.27% accuracy with 20 epochs. It identified leopards, hogs, birds, and genet civets with over 90% accuracy but struggled with the blank class. Training was completed in under eight hours on a single laptop, showcasing the efficiency of lightweight models. Our findings underscore the potential of computer vision in conservation, enabling rapid and accurate analysis of large datasets that would take large amounts of time and workers to go through manually. This work highlights the importance of hyperparameter tuning in enhancing model performance, paving the way for more effective automated wildlife monitoring tools. Future work will focus on improving accuracy for challenging classes and testing model adaptability in different environments.