An automatic identification method of common species based on ensemble learning

Abstract

Camera traps are an important tool for animal resource surveys, allowing non-invasive wildlife image capture and providing essential data for species identification. However, the vast number of images generated requires significant manual effort for sorting, limiting its development in biodiversity studies. Deep learning offers a promising solution by accurately identifying species from large datasets, enhancing processing efficiency, and reducing costs. While existing deep learning methods have achieved significant success in species identification, they often struggle with accurately recognizing all species due to the class imbalance prevalent in camera trap datasets, which limits the application of deep learning models in biodiversity monitoring. This study proposed an ensemble learning method based on common species modeling to automatically identify common species, which constitute the majority of camera trap datasets. We utilized three base models: ResNet-18, ResNeXt-50, and ViT-Base to validate our method on the Snapshot Serengeti dataset. The experimental results showed that the performance of the ensemble learning method improved with the performance of the selected base model. When ResNeXt-50 was used as the base model, the recall and precision of all common species on the in-sample test set exceeded 98 % and 97 %, respectively, except for Gazelle Grants. The automation rate of the ensemble model was 80.67 %, and the omission error of rare species was 2.03 %. On the out-of-sample test set, all species except for Zebra, Buffalo, and Gazelle Grants had a recall of over 95 %. Apart from Gazelle Grants, the precision for the other species was above 90 %. The automation rate of the ensemble model was 72.27 %, and the omission error of rare species was 5.31 %. Our method achieved the automatic identification of common species, thus reducing the workload of manual sorting. In addition, our approach separated rare species images from the dataset by identifying common species, minimizing potential omission errors. As a result, ecologists focusing on rare species only need to handle rare species images that only represent a small proportion of the dataset.