AAI_2025_Capstone_Chronicles_Combined

While Figure 1 shows slight variation in class distribution between in-location and out-location datasets, monthly stratification in Figures 3 and 4 reveals a substantial difference in pattern. Temporal shift is also observed for both datasets. Despite this imbalance, the EDA phase confirmed that the dataset is clean, well-structured, and appropriate for simulating real-world deployment scenarios. It also highlighted challenges such as data sparsity for rarer species, spatio-temporal shifts, and the need for further techniques like data augmentation to address these issues in future iterations. Camera trap monitoring has become an essential tool for wildlife conservation and biodiversity assessment, with millions of images captured annually from trail cameras deployed worldwide (Morris, 2023). Machine learning (ML) has revolutionized the processing of these images, enabling automated animal detection, species classification, and the filtering of empty images (Dolan et al., 2020). However, real-world deployments present challenges, such as adapting models to different ecosystems, camera configurations, and new species, requiring continuous improvement pipelines to maintain model performance (Zhong et al., 2023). For wildlife image classification, the typical workflow involves first using object detection to localize animals in the images, followed by a classification model to identify the species. One prominent tool in this space is MegaDetector, a general-purpose object detection model that excels in camera trap applications. Trained on millions of images from diverse ecosystems, MegaDetector detects animals, humans, and vehicles with high accuracy, while also filtering out empty images (Microsoft, n.d.; Leorna & Brinkman, 2022). Despite its strengths, MegaDetector does not provide species-level classification, which is where our system complements it by integrating CNNs for classification. To improve species identification, we used transfer learning with CNNs, fine-tuning pre-trained models like ResNet-18 on the CCT20 dataset. This approach allows the model to focus on animals rather than the full image context, improving classification accuracy across species, including rare ones. Additionally, we Background Study

284