An automatic method for removing empty camera trap images using ensemble learning.

Area of Science:

• Ecology and Conservation Technology
• Artificial Intelligence in Wildlife Monitoring

Background:

• Camera traps generate vast amounts of image data, with many images being empty, posing a significant data management challenge.
• Current deep learning methods for identifying empty images require extensive labeled datasets, incurring high time and labor costs.
• Small datasets often lead to deep learning models with high omission errors, risking the loss of valuable animal presence data.

Purpose of the Study:

• To develop an efficient deep learning approach for automatically identifying and removing empty camera trap images using small datasets.
• To minimize the omission error of animal images while reducing manual data processing costs.
• To provide users with customizable schemes for empty image removal based on acceptable omission error rates.

Main Methods:

• An ensemble learning approach based on conservative strategies was developed for deep convolutional neural networks (DCNNs).
• Three distinct automatic identification schemes were proposed, allowing users to select based on tolerance for omission errors.
• The method was evaluated on a small-sized dataset of camera trap images.

Main Results:

• The proposed schemes automatically identified and removed 50.78%, 58.48%, and 77.51% of empty images with omission errors of 0.70%, 1.13%, and 2.54%, respectively.
• The approach successfully avoided omitting species information, with only minor alterations to species occurrence frequencies.
• Significant reductions in time and personnel costs associated with manual image review were achieved.

Conclusions:

• The developed ensemble learning approach offers a viable solution for processing large camera trap datasets with limited labeled data.
• It provides a cost-effective alternative to manual review, crucial for ecological research and species monitoring.
• The customizable schemes ensure flexibility for researchers managing different levels of acceptable omission errors.