An Adaptive Automatic Approach to Filtering Empty Images from Camera Traps Using a Deep Learning Model

Camera traps are widely used in wildlife surveys because they are non-invasive, low-cost, and highly efficient. Camera traps deployed in the wild often produce large datasets, making it increasingly difficult to manually classify images. Deep learning is a machine learning method that provides a tool to automatically identify images, but it requires labeled training samples and high-performance servers with multiple Graphics Processing Units (GPUs). However, manually preparing large-scale training images for training deep learning models is labor intensive, and the high-performance servers with multiple GPUs are often not available for wildlife management agencies and field researchers. Our study explores an adaptive deep learning method to use small-scale training sets and a commonly-available, desktop personal computer (PC) to achieve automatic filtering of empty camera images. Our results showed that by using 29,192 training samples, the overall error, commission error, and omission error of the proposed method on a PC were 2.69%, 6.82%, and 6.45%, respectively. Moreover, the accuracy of our method can be adaptively improved on PCs in actual ecological monitoring projects, which would benefit researchers in field settings when only a PC is available. (c) 2021 The Wildlife Society.