The successes and pitfalls: Deep-learning effectiveness in a Chernobyl field camera trap application

Area of Science:

• Ecology
• Artificial Intelligence
• Wildlife Conservation

Background:

• Camera traps are crucial for wildlife monitoring, providing data on animal populations and ecosystem health, especially in challenging environments.
• Manual image processing is time-consuming and expensive, leading to the adoption of machine learning, particularly convolutional neural networks (CNNs).
• A significant challenge for CNNs is the need for vast datasets (millions of images) to achieve high accuracy, which is often impractical.

Purpose of the Study:

• To investigate how specific camera trap placement factors influence the accuracy of a deep learning model trained on a limited dataset.
• To assess the impact of environmental variables like weather and daylight on the performance of CNNs in classifying camera trap imagery.
• To explore methods for improving CNN accuracy in wildlife monitoring, even with smaller training datasets.

Main Methods:

• A convolutional neural network (CNN) was transfer-trained to identify 16 object classes (14 animal species, humans, fires) using 9,576 camera trap images from the Chernobyl Exclusion Zone.
• The study analyzed the correlation between CNN classification success and environmental factors including wind speed, cloud cover, temperature, image contrast, and precipitation.
• The model's performance was evaluated under different daylight conditions and in the presence or absence of precipitation.

Main Results:

• No significant correlation was found between CNN success rates and most ambient environmental conditions (wind speed, cloud cover, contrast).
• A potential positive relationship was observed between ambient temperature and CNN model success.
• The CNN model demonstrated higher accuracy during daylight hours and in the absence of precipitation.

Conclusions:

• Environmental factors like weather and lighting can introduce false negatives and positives, challenging automated image classification.
• While some environmental conditions did not significantly impact CNN performance, temperature showed a possible positive correlation.
• Training deep learning models with dynamic datasets that account for ambient conditions can minimize their impact on classification accuracy for wildlife monitoring.