Understanding External Influences on Target Detection and Classification Using Camera Trap Images and Machine...

Area of Science:

• Ecology
• Computer Science
• Conservation Biology

Background:

• Automating camera trap (CT) image processing with machine learning (ML) is crucial for time-sensitive conservation efforts.
• Factors influencing the accuracy of ML-based CT image analysis remain largely uninvestigated.

Purpose of the Study:

• To evaluate the impact of various environmental and subject-related factors on wildlife and human detection and classification in CT images.
• To compare the performance of human analysts versus ML algorithms in processing CT images.
• To identify key challenges and potential improvements for ML-based CT image analysis.

Main Methods:

• Analysis of CT images from western Tanzania, including wildlife data from existing datasets and human data from experimental setups.
• Evaluation of both human analyst and ML approaches at the detection and classification levels.
• Assessment of factors such as occlusion, distance, vegetation, size, height, orientation, time of day, and color.

Main Results:

• Distance and occlusion, along with dense vegetation, significantly affected detection probability (DP) and correct classification (CC).
• Human analysts achieved higher DP (81.1%) and CC (76.6%) for wildlife compared to ML (41.1% DP, 47.5% CC).
• Both methods showed comparable performance in detecting humans during daylight and dusk.

Conclusions:

• Human analysts currently provide superior accuracy for wildlife detection and classification in CT images over ML.
• Environmental factors like distance and vegetation density are critical limitations for current ML-based CT image processing.
• Adherence to recommendations is expected to improve ML performance, making it a valuable tool for biodiversity conservation and threat monitoring.