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

AI summary

Overview

• Research area: Applied computer vision for ecology and conservation monitoring; explainable AI (XAI) for object detection.
• Technical level: Intermediate. Familiarity with object detection and saliency/attribution concepts helps, but the paper defines its metrics plainly.
• Scope: A case study that trains a Faster R-CNN seal detector on aerial drone imagery from Glacier Bay National Park and evaluates three families of post-hoc explanations (gradient-based CAM, LIME, and perturbation-based deletion tests) against field-relevant criteria.

What This Paper Is About

Deep neural detectors are the most accurate tools available for wildlife monitoring, but ecologists are hesitant to rely on them because their reasoning is opaque, and a single false positive or negative has real operational and ethical costs. Using harbor seal detections in aerial drone imagery, the authors ask whether post-hoc explanations can supply evidence for individual predictions, expose systematic failure modes, and make model outputs auditable enough for conservation decisions. Their framing is that explainability should be routine practice when validating ecological computer vision models, not an afterthought.

Key Contributions

1. A full explainability pipeline for ecological object detection. The authors pair a Faster R-CNN detector with HiResCAM, LayerCAM, LIME, and perturbation-based (deletion) explanations, and release code at https://github.com/duke-trust-lab/on-thin-ice.
2. Three ecologically motivated evaluation axes. They define localization fidelity (attribution ratio and maximum saliency hit-rate), faithfulness (flip rate and confidence drop under mask, noise, and blur perturbations), and diagnostic utility (manual review of false positives with saliency overlays).
3. An adapted LIME for multi-instance detection. They contribute improved superpixel segmentation in LAB color space with SLIC, a detector-aware weighting scheme (confidence, area, or uniform modes), and proximity-based spatial suppression for cleaner attribution maps.
4. Concrete documentation of failure modes and next steps. The analysis identifies black ice and rocks as recurrent confusions and translates this into recommendations such as targeted data curation and augmentation with negative examples.

Main Findings

• Detections are grounded in the animals, not just context. Across 76 test images, LayerCAM achieved an attribution ratio of 67.69 ± 25.13% with a maximum-saliency hit rate of 94.70% (125/132 boxes). HiResCAM reached 58.56 ± 30.83% attribution ratio and 69.70% hit rate (92/132 boxes). LIME reached 48.76 ± 29.51% attribution ratio and 92.42% hit rate (122/132 boxes).
• Removing the seal reliably removes the detection. Masking produced a mean confidence decrease of 0.97 with a 97.33% flip rate; additive noise produced 0.87 with an 80.00% flip rate; blur produced only 0.27 with a 21.33% flip rate.
• Blur is the weakest perturbation. When predictions did not flip, confidence fell by only 0.07 under blur, versus 0.10 for mask and 0.37 for noise, indicating the model can sustain detections from coarse morphology and intact body contours.
• Illustrative case: a seal detected at full confidence (1.00) is driven to no detection by masking or noise, while blur only reduces confidence to 0.77.
• Black ice and rocks are the dominant confusion. Attribution maps consistently highlighted black ice regions for a false positive example (0.58 confidence), showing the model keys on low-level visual similarity rather than biological features. Other errors included one missed seal annotation, four dark shapes at image edges, one dark shape in open water, and one merged detection of two adjacent seals.
• Color cues are not essential. Black masking (RGB 0,0,0) achieved a 100% flip rate on both models tested, exceeding mean background masking (97.33%), noise (80.00%), and blur (21.33%), suggesting decisions depend more on structure and spatial features than on color.
• Black masking required very small edits. YOLO produced successful counterfactuals on 44 of 76 images (57.9%) with an average of 1.2 ± 0.4 superpixels and 0.45% ± 0.20% of image area changed, with a mean confidence drop of 0.73 ± 0.15. Faster R-CNN produced successful counterfactuals on 40 of 76 images (52.6%), averaging 1.3 ± 0.5 superpixels and 0.52% ± 0.25% of area, with a mean confidence drop of 0.75 ± 0.18.
• Detector performance. The best model was obtained at epoch 67 with validation mAP (IoU 0.5 to 0.95) of 0.61, mAP50 of 0.95, and mAR of 0.69, with training and validation losses of 0.11 and 0.13. Section 3.1 reports a final mAP50 of 0.95 and mAR of 0.69 on the held-out test set, while Appendix B reports test-set mAP of 0.65, mAP50 of 0.98, mAP75 of 0.78, and mAR100 of 0.72, with size-stratified mAP of 0.62 for small, 0.65 for medium, and 0.77 for large objects.
• Explanations are method-sensitive. The authors state that all explainers are post-hoc approximations — CAM is sensitive to layer choice and upsampling, LIME to superpixel segmentation and perturbation kernel — and that deletion edits can induce distribution shift. Faithfulness is measured via post-NMS confidence, which may be miscalibrated, and localization is evaluated against bounding boxes rather than pixel masks.

Methodology in Plain English

The team collected 1,974 aerial drone images of harbor seals from Glacier Bay National Park using a Wingtra One Gen II fixed-wing VTOL platform with a Sony Alpha 6100 APS-C camera and a Sony E 20mm f/2.8 lens, flying at 60–85 m altitude and 9–22 m/s airspeed under NOAA and NPS permits. Images were tiled to 640×640 pixels and manually annotated with Labelme. The dataset was split into 1,744 training, 151 validation, and 76 test images. They fine-tuned a Faster R-CNN with a ResNet-50 backbone and Feature Pyramid Network, initialized from COCO weights, using geometric and color augmentation on training data only, trained with stochastic gradient descent, gradient clipping, warmup scheduling, and early stopping (batch size 8, learning rate 0.0090, momentum 0.874, weight decay 0.0001, selected by grid search).

To explain the detector, they applied three approaches. Gradient-based CAM variants (HiResCAM, LayerCAM) use element-wise activation–gradient products at the final convolutional block (backbone.body.layer4) rather than global pooling, producing a heatmap of influential regions. Their adapted LIME segments images into superpixels, masks them, and fits a local surrogate to score each region's importance for multiple detected instances at once. Perturbation-based explanations greedily edit the most damaging superpixels inside or near a bounding box until confidence falls below a threshold of 0.5 or 80 iterations are reached, using mask, Gaussian blur (σ = 5), or additive uniform noise; detections are matched by class and IoU with a 0.2 threshold. Explanations were then scored by attribution ratio, maximum-saliency hit-rate, flip rate, and confidence drop, and false positives were manually reviewed.

Why This Matters

Impact on research. The paper argues that accuracy alone is insufficient for ecological deployment and offers a reusable evaluation recipe — localization fidelity, faithfulness, diagnostic utility — that other conservation computer vision projects can apply. It also highlights a gap: most XAI work targets classification, and the authors state that to their knowledge no prior studies evaluate explanations against ecologically relevant criteria in conservation workflows.

Real-world applications:

• Aerial monitoring of pinnipeds and other wildlife in remote, ice-covered habitats where surveys are expensive and infrequent.
• Triage of automated detections, giving reviewers a ranked set of characteristics to check (for example, dark shapes at image edges or black ice patches) rather than re-inspecting every frame.
• Data curation and augmentation planning, by using explanations to identify underrepresented negative examples such as black ice and rock.
• Model thresholding and trust calibration, where practitioners decide when an automated detection is reliable enough to act on.

Industry relevance. Any domain deploying detectors in high-stakes or regulated settings — environmental compliance, infrastructure inspection, remote sensing analytics — faces the same auditability demand. The paper's finding that attribution methods converge on the same confounding structures when a model fails is directly transferable to failure analysis pipelines, and the public code release lowers the barrier to adopting these audits.

Future Directions

• Targeted data curation and augmentation. The authors recommend adding more negative examples of black ice and refining the training set to better capture background variability, based on the observed confusion.
• Generalization testing beyond the current site. The authors explicitly note they do not claim habitat- or model-general conclusions, leaving open whether these results transfer to other ice and rock environments.
• Better faithfulness measurement. Because faithfulness was measured via post-NMS confidence, which may be miscalibrated and need not correspond to error rates or mAP changes, alternative evaluation approaches remain an open question.
• Pixel-level localization evaluation. Localization was scored against bounding boxes rather than pixel masks, which may penalize explanations that focus on salient subregions or spill slightly outside labels, motivating mask-based benchmarks.

Target Audience

Ecologists and conservation practitioners who want to understand what an automated detector is actually reacting to before trusting its output; computer vision researchers working on explainability for object detection rather than classification; and applied machine learning engineers in environmental monitoring, remote sensing, or any domain where model decisions must be defended to regulators or stakeholders. Readers seeking a pure benchmark or a new architecture will not find one here — this is a case study about validation practice.

Authors’ 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.