Det-LIME: Detector-Aware, Multi-Instance Local Interpretable Model-Agnostic Explanations for Automated Marine Mammal Detection
Det-LIME is a detector-aware, multi-instance adaptation of LIME that produces instance-specific, box-aligned explanations for object detection. It was evaluated on aerial drone imagery for harbor seal detection, with an additional seabird case study to assess generality, and compared with vanilla LIME and Stabilized LIME.
Researchers proposed Det-LIME, a detector-aware, multi-instance adaptation of Local Interpretable Model-Agnostic Explanations (LIME) for object detection in marine mammal monitoring. The method combines per-detection weighting, a proximity kernel emphasizing regions near each bounding box, and Intersection-over-Union-based matching to track instances across perturbations. Evaluation on aerial drone imagery for harbor seal detection and a seabird case study showed improved instance-specific, box-aligned explanations compared to vanilla LIME and Stabilized LIME.
Det-LIME addresses limitations of classification-oriented explainability in detection tasks by generating box-aligned, instance-specific explanations. Key technical components include per-detection weighting, a proximity kernel focused on each bounding box, and IoU-based matching to maintain instance identity across perturbations. This approach enables higher-resolution, biologically relevant visual explanations for multi-instance scenes.
The integration of explainability into conservation workflows remains limited. Det-LIME's detector-aware design could support debugging, targeted data augmentation, and refined data collection in marine mammal monitoring, potentially increasing adoption of black-box detectors in ecological research.
Improved explainability for object detectors can reduce debugging time, improve model trust, and enable targeted data collection in conservation and ecological monitoring. This may lower barriers to deploying AI in environmental applications and support regulatory or stakeholder communication.
Next observable signals include application of Det-LIME to other detection architectures and species, publication of code or benchmarks, and adoption by conservation technology platforms. Further validation on diverse imagery and integration into monitoring pipelines would indicate practical impact.