AAI_2025_Capstone_Chronicles_Combined

explored incorporating temporal features, such as the date the image was captured, which can expose seasonal and daily activity patterns of animals. Temporal encoding has been shown to improve classification performance by capturing these patterns. Modern camera trap ML systems increasingly rely on confidence-based filtering to maintain high accuracy while maximizing automated processing (Le Coz et al., 2024; Yang et al., 2021). Research has shown that setting confidence thresholds at 95% or higher can achieve over 99% accuracy on high-confidence predictions, though this comes at the cost of reduced coverage (Yang et al., 2021; Leorna & Brinkman, 2022). The development of calibration techniques for deep learning models in camera trap applications has become crucial for reliable deployment. Proper calibration ensures that confidence scores accurately reflect the likelihood of correct predictions, enabling more informed decisions about when to accept automated classifications versus triggering human review. Böhner et al. proposed a comprehensive semi-automated workflow specifically designed for long-term monitoring programs. This workflow includes automated image classification, quality checking of predictions, model retraining capabilities, and manual review processes for uncertain classifications. The approach emphasizes the importance of continuous model updating as monitoring programs evolve and encounter new species or environmental conditions (Böhner et al., 2023; Liu et al., 2024). Active learning has emerged as a powerful strategy for minimizing human annotation effort while maintaining model performance. Norouzzadeh et al. demonstrated that active learning systems can match state-of-the-art accuracy with a 99.5% reduction in manual labeling effort, using only 14,100 manually labeled images instead of millions (Norouzzadeh et al., 2019; Auer et al., 2020). This approach is particularly valuable for camera trap applications, where obtaining labeled data is expensive and time-intensive. Our project aims to combine both strategies for human interaction. A dashboard display containing unsupervised metrics and key comparisons with learned data will be generated for further user

285