On Thin Ice: Towards Explainable Conservation Monitoring via Attribution and Perturbations

TL;DR

This work tackles the trust gap in ecological computer vision by applying post-hoc explanations to a harbor-seal detector trained on aerial Glacier Bay imagery. By combining HiResCAM, LayerCAM, LIME, and perturbation-based tests, it evaluates localization fidelity, faithfulness, and diagnostic utility, revealing that explanations align with seal contours but also expose confounds such as black ice. The results demonstrate that explanations provide auditable evidence for predictions and identify concrete failure modes, guiding data curation and model improvements. Overall, the proposed workflow moves conservation monitoring toward auditable, decision-support tools rather than opaque predictions, with practical implications for field deployment and model validation.

Abstract

Computer vision can accelerate ecological research and conservation monitoring, yet adoption in ecology lags in part because of a lack of trust in black-box neural-network-based models. We seek to address this challenge by applying post-hoc explanations to provide evidence for predictions and document limitations that are important to field deployment. Using aerial imagery from Glacier Bay National Park, we train a Faster R-CNN to detect pinnipeds (harbor seals) and generate explanations via gradient-based class activation mapping (HiResCAM, LayerCAM), local interpretable model-agnostic explanations (LIME), and perturbation-based explanations. We assess explanations along three axes relevant to field use: (i) localization fidelity: whether high-attribution regions coincide with the animal rather than background context; (ii) faithfulness: whether deletion/insertion tests produce changes in detector confidence; and (iii) diagnostic utility: whether explanations reveal systematic failure modes. Explanations concentrate on seal torsos and contours rather than surrounding ice/rock, and removal of the seals reduces detection confidence, providing model-evidence for true positives. The analysis also uncovers recurrent error sources, including confusion between seals and black ice and rocks. We translate these findings into actionable next steps for model development, including more targeted data curation and augmentation. By pairing object detection with post-hoc explainability, we can move beyond "black-box" predictions toward auditable, decision-supporting tools for conservation monitoring.

Figures (2)

• Figure 1: Visual explanations of harbor seal detections using Faster R-CNN. Top row: Successful detection. In the attribution maps, a color spectrum represents relative importance, with warm colors (e.g., red, orange, yellow) indicating high contribution and cool colors (e.g., blue) indicating little or no contribution. (a) Original detections with bounding boxes. (b) LayerCAM attribution (c) HiResCAM attribution (d) LIME attribution. Bottom row: Challenging failure case where black ice is misidentified as a seal. (e) Original detections with bounding boxes, including a false positive with 0.58 confidence. (f) LayerCAM attribution (g) HiResCAM attribution (h) LIME attribution. The attribution maps reveal that the model confuses black ice with seals, indicating a limitation in generalizing to environments with black ice or rocks.
• Figure 2: Perturbations of a successful seal detection. (a) Original detection with confidence score 1.00. (b) Masking the seal body leads to no detection. (c) Adding noise leads to no detection. (d) Blurring reduces confidence to 0.77 but preserves detection. These perturbations demonstrate how the model relies on seal contours and feature coherence for reliable predictions.