Source-linked AI summary
Det-LIME: Detector-Aware, Multi-Instance Local Interpretable Model-Agnostic Explanations for Automated Marine Mammal Detection
Jiayi Zhou, David W. Johnston, Brinnae Bent
TL;DR
Black-box ecological detectors need explanations that distinguish evidence for multiple animals rather than collapsing detections into a single image-level attribution. Det-LIME extends LIME with instance weighting, box-focused proximity, and IoU matching; evaluations reported improved multi-instance attribution and actionable diagnostic insights, while limitations include superpixel sensitivity, bounding-box-based fidelity assessment, noisy labels, and the absence of component ablations.
Problem
Classification-oriented explainability methods often blur evidence across multiple detections in ecological scenes, limiting instance-level interpretation for social or colonial organisms.
Method
Det-LIME is a detector-aware, model-agnostic LIME adaptation that uses per-detection weighting, a proximity kernel, and IoU matching to produce instance-specific explanations.
Results
Det-LIME produced box-conditioned, instance-stable attributions that localized valid seal evidence and exposed background-driven false detections, improving multi-instance attribution.
Takeaways & Limitations
The explanations supported confidence assessment, debugging, targeted data augmentation, model refinement, and more focused ecological data collection.
Takeaways & Limitations
The evaluation was limited by superpixel and perturbation sensitivity, bounding-box-based fidelity measures, noisy labels, and the absence of systematic component ablations.
Abstract
from arXiv · showhide
Despite the rapid uptake of black-box object detectors in marine mammal research and monitoring, explainability techniques are rarely integrated into conservation workflows. Furthermore, most classification-oriented explainability tools are ill-suited to detection tasks involving imagery of social organisms or those with colonial life histories, as they ignore multiple detections within a scene and produce single-instance outputs that blur evidence across individuals. These methods also generate low-resolution, often biologically irrelevant visuals, limiting their utility for debugging, targeted data augmentation, and refined data collection. We proposed Det-LIME, a detector-aware, multi-instance adaptation of Local Interpretable Model-Agnostic Explanations (LIME) that produced instance-specific, box-aligned explanations by combining per-detection weighting, a proximity kernel that emphasizes regions near each box, and Intersection-over-Union-based matching to track the same instance across perturbations. We evaluated Det-LIME on aerial drone imagery for harbor seal detection, with an additional seabird case study to assess generality, and compared it with vanilla LIME, Stabilized LIME, Deterministic LIME, and gradient-based attribution methods. Using the Attribution Ratio and Max Saliency Hit Rate metrics, we showed that Det-LIME consistently improved multi-instance attribution. In practice, these higher-resolution, instance-aware explanations provide insight into model outputs and support post-processing, debugging, and actionable improvements in modeling and data collection or augmentation.
1 Introduction
Ecological computer-vision workflows increasingly depend on black-box detectors, creating a need for explanations that support trustworthy, operationally relevant decisions. Det-LIME addresses this gap by producing instance-specific, box-aligned evidence for multiple detections in an image.
- Motivation: Computer vision supports scalable biodiversity monitoring by automatically identifying, counting, and locating species and landscape features in large image and video datasets.
- Motivation: Black-box detectors create a need for evidence that predictions rely on animal morphology, reveal failure modes, and support defensible operational thresholds.
- Motivation: Classification-oriented explanations are poorly suited to crowded ecological images because detectors produce multiple localized predictions rather than one image-level decision.
- Contribution: Det-LIME adapts LIME for detection by weighting instances, emphasizing regions near each bounding box, and matching detections across perturbations with IoU.
- Contribution: The method targets higher-resolution, per-animal evidence aligned with how ecologists review detections and make policy-relevant decisions.
2 Related Work
Ecological XAI remains less developed than in several other application areas, while existing detection explanations often lack model generality or instance-level precision. Det-LIME extends LIME to provide spatially specific explanations for multiple objects in ecological imagery.
- Explainability in Ecology: Ecological explainable AI remains comparatively underdeveloped despite recognized needs in high-stakes conservation contexts.
- Explainability in Ecology: LIME and SHAP have supported localized or feature-based explanations in ecological classification, species distribution, and environmental modeling tasks.
- Detection Explainability: Grad-CAM and saliency methods can explain individual detections but typically require access to model architecture or internal feature maps.
- Detection Explainability: Conventional LIME is model-agnostic but generally explains only one detected instance, limiting its use for multi-object ecological imagery.
- Det-LIME: Det-LIME preserves LIME’s model-agnostic approach while extending local explanations to instance-level, spatially specific detection outputs.
3 Methods
Det-LIME extends LIME for multi-instance object detection by tracking, weighting, and spatially focusing individual detections. Its pipeline produces box-aligned pixel-level explanations while evaluating attribution against detector outputs.
- Image segmentation: Enhanced SLIC segmentation enforces compactness, minimum-size, and connectivity constraints, with optional filtering of visually insignificant segments.The filtering step removes regions such as uniformly dark segments to reduce background noise and improve explanation fidelity.
- Det-LIME adaptation: Det-LIME extends LIME to multi-instance detection by explicitly selecting target instances, weighting them, scoring perturbations, and fitting a local linear explanation.The method is designed for images containing multiple detections rather than a single image-level prediction.
- Target instance identification and weighting: The detector defines target instances as sufficiently confident detections of the selected class, each represented by a bounding box, class label, and confidence.Target detections are collected into a set, and the framework supports confidence-based, spatial-extent, or equal weighting.
- Multi-instance perturbation and scoring: The method masks subsets of superpixels, aggregates contributions from all target instances, and uses the resulting scalar score as LIME’s response variable.This score reflects how well the target detection set is preserved under each perturbation.
- Multi-instance perturbation and scoring: Perturbed detections are matched to original targets using IoU, while an overlap threshold excludes detections with minimal spatial correspondence.Matched confidence values above 0.95 are capped at 0.95 for scoring, preventing a single highly confident prediction from dominating the instance-level score.
- LIME linear model fitting: A weighted linear surrogate assigns superpixel coefficients that represent their contribution to preserving the aggregated multi-instance score.Because coefficients correspond to individual superpixels, their estimated contributions can be mapped back to image regions.
- Spatial focusing and pixel-level explanation: Spatial focusing combines superpixel importance, overlap with each instance, proximity to its center, and instance weights into a normalized pixel-level saliency map.The inner and outer maximizations select the most relevant superpixel and instance contributions, respectively; black or filtered segments may be excluded and Gaussian smoothing applied.
- Evaluation protocol: The evaluation compares LIME, DLIME, SLIME, and Det-LIME using attribution metrics aligned with ground-truth object regions rather than standalone detector quality.Det-LIME explains all relevant detections, whereas the other methods generate explanations only for the single highest-confidence object; evaluations report valid matched instances and include an adjusted one-to-one comparison.
4 Results
Det-LIME substantially improved instance-level attribution across harbor seal and penguin detection tasks, while visual comparisons showed sharper, more coherent explanations than baseline methods. These results support its use for actionable interpretation in ecological multi-instance detection.
- Det-LIME substantially improved Attribution Ratio and Max Saliency Hit Rate across harbor seal and penguin tasks, exceeding strongest baselines by more than 20–30 percentage points in several comparisons.Table 1 averaged metrics over matched detection instances rather than images.
- Bounding-box-aligned perturbations and IoU-based matching limited attribution drift, preserved correspondence with individual objects, and reduced spurious background signals.
- Det-LIME produced sharper, more coherent attributions aligned with annotated animal regions, whereas baseline maps often highlighted ice, rock, or water textures.The visual difference was especially pronounced for small, densely clustered penguins, where coarse perturbations posed a severe challenge.
- Det-LIME’s instance-aware explanations better supported threshold setting, post-processing, and targeted ecological data collection.
5 Discussion
Det-LIME provides instance-specific visual evidence that clarifies both valid detection cues and recurring failure modes in ecological object detection. Its diagnostics support interpretation and model improvement, while its evaluation remains constrained by superpixel sensitivity, bounding-box-based metrics, non-isolated components, and unassessed deployment factors.
- Correct detections: For correct seal detections, Det-LIME emphasized superpixels overlapping the animal or its boundary, whereas conventional LIME methods often merged animals with adjacent background.This contrast helps distinguish evidence associated with the target from contextual evidence.
- Failure diagnosis: Failure cases showed Det-LIME identifying dark ice textures, co-occurring albatrosses, and habitat boundaries as visual cues behind false-positive seal and penguin detections.Compared with bounding boxes alone, the attribution maps exposed the image regions influencing each erroneous prediction.
- Interpretation: Det-LIME heatmaps indicate model behavior rather than animal anatomy, so strong attribution on nearby background may reflect contextual information combined with animal features.The method perturbs superpixel combinations and measures effects on a particular detection, but does not directly reveal internal feature relationships.
- Attribution performance: Det-LIME outperformed vanilla LIME and variants SLIME and DLIME on multi-instance attribution, concentrating saliency within queried detections and anchoring peak attribution to the same instance.The evaluation used Attribution Ratio and Max Saliency Hit Rate to assess instance fidelity.
- Practical implications: The approach can guide targeted negative examples, habitat-diverse training data, model refinement, and cautious use of automated conservation predictions.These applications follow from recurring errors involving ice, rock, albatrosses, and habitat boundaries.
- Limitations: Interpretation is limited by superpixel and perturbation-kernel sensitivity, bounding-box-based fidelity measures, noisy labels, and evaluation of the complete pipeline without component ablations.Runtime efficiency, operator burden, and field usability were also not assessed.
Note on references
The paper situates several cited contributions within computer science and machine-learning publication practices, where conference and workshop proceedings often serve as archival venues.
- Conference and workshop proceedings are important publication venues for foundational computer-science and machine-learning contributions.
- Several references in this work therefore appear in conference proceedings that function as primary archival records.
- The reference note explains this venue choice rather than presenting a methodological or empirical finding.
A.1 Harbor Seal Data
The harbor seal dataset was collected through fixed-wing drone surveys over historically sampled regions and prepared by filtering, tiling, and manually annotating seal-containing imagery.
- A.1 Harbor Seal Data: Drone surveys used a Wingtra One Gen II fixed-wing platform with a Sony Alpha 6100 camera and flew at approximately 60–85 m altitude.
- A.1 Harbor Seal Data: Flights followed parallel transects through glacial fjords, concentrating coverage in inner regions where seal densities were highest.
- A.1 Harbor Seal Data: Training images were filtered to retain photographs containing at least one harbor seal on floating ice, reducing potential negative bias.
- A.1 Harbor Seal Data: The retained images were divided into 640 × 640-pixel tiles and manually inspected for annotation of all harbor seal locations.
B.1 Harbor Seal Detection Task using Faster R-CNN
The harbor seal detector used a fully fine-tuned Faster R-CNN, augmented training data, and held-out testing; validation peaked at epoch 67 with reported mAP values of 0.61 and 0.95 at IoU 0.5.
- B.1 Harbor Seal Detection Task using Faster R-CNN: The detector was Faster R-CNN with a ResNet-50 backbone and Feature Pyramid Network, initialized on COCO and fully fine-tuned for seal detection.
- B.1 Harbor Seal Detection Task using Faster R-CNN: Training used geometric and color augmentation, while validation received only normalization and the test set remained fully held out without augmentation.
- B.1 Harbor Seal Detection Task using Faster R-CNN: Optimization used stochastic gradient descent with gradient clipping, warmup scheduling, early stopping, and experiment tracking.
- B.1 Harbor Seal Detection Task using Faster R-CNN: At epoch 67, validation performance peaked at mAP 0.61 and mAP50 0.95 after hyperparameter tuning and training with early stopping.The reported mAP averages IoU thresholds from 0.5 to 0.95.
B.2 Penguin Detection Task using YOLOv9
The penguin detection case study used YOLOv9c on reformatted training, validation, and test datasets with augmentation and mixup regularization to improve generalization.
- B.2 Penguin Detection Task using YOLOv9: The penguin detector used the YOLOv9c architecture on 2,644 training, 288 validation, and 321 test images.
- B.2 Penguin Detection Task using YOLOv9: Images were reformatted into YOLOv9 format for consistency across the dataset splits.
- B.2 Penguin Detection Task using YOLOv9: Training applied rotation, translation, scaling, horizontal flips, color adjustments, and mixup regularization to improve generalization and reduce overfitting.