AAI_2025_Capstone_Chronicles_Combined

reduce susceptibility to misleading background patterns. Context-aware cropping in the detection stage may also help minimize irrelevant background features before classification.

Overall, these findings validate the effectiveness of a fully custom, from-scratch convolutional architecture on a challenging wildlife classification task. Beyond establishing competitive accuracy, the analysis highlights clear and actionable directions for improvement, positioning ScratchResNet as both a strong research contribution and a practical foundation for future experimentation.

For pre-deployment evaluation, we assessed each model on two key dimensions: multi-class classification and confidence reliability.

Figure 9 presents the multi ‑ class classification results for both in ‑ location and out ‑ location test sets from the first year of evaluation. Both ResNet ‑ 18 variants (Models 1 and 2) outperform the baseline Model 0. However, the inclusion of temporal vectors in Model 2 did not lead to any appreciable improvement in classification performance over Model 1.

291