AAI_2025_Capstone_Chronicles_Combined

analysis. New images with low prediction confidence rates will be labeled, and human annotators will be involved to increase the dataset’s robustness.

​ Camera trap applications face unique challenges in incorporating new species and adapting to changing conditions without catastrophic forgetting. Class incremental learning approaches address these challenges by maintaining performance on previously learned species while incorporating new ones through careful exemplar management and feature space organization (Zhu et al., 2022). The adaptive exemplar assignment strategy demonstrates how to balance limited memory constraints with the need to retain knowledge of rare species, achieving 77% accuracy using only 4% of original species data. This approach is particularly relevant for long-term monitoring programs where new species may be encountered over time. ​ Fine-tuning approaches using pre-trained models have shown significant success in camera trap applications. EfficientNetV2-L models fine-tuned on camera trap data achieved 88.8% accuracy with careful attention to data augmentation and overfitting prevention (Doan & Le-Thi, 2023). These approaches demonstrate the importance of leveraging large-scale pre-training while adapting to the specific challenges of wildlife imagery. Model compression techniques enable the deployment of sophisticated models on resource-constrained edge devices, achieving lightweight recognition models with only 4.73% accuracy loss compared to full-scale models (Zhong et al., 2023). This capability is essential for real-time field deployment and continuous monitoring applications, ensuring that models remain efficient and scalable.

Experimental Methods

The WildScan project employs two variants of the ResNet ‑ 18 architecture, each adapted to the characteristics of the dataset and the classification task. To realistically assess performance over time, the dataset is partitioned both spatially (by location) and temporally (by year). For baseline training, we used the entire first year of data to ensure the models were exposed to all species that might appear in

286