Camera Trap Distance Sampling (CTDS) enables population density estimation of unmarked species from camera trap data by incorporating animal-to-camera distances. In this study, we systematically evaluate the sensitivity of CTDS-derived density estimates to variation in key pipeline parameters, using a high-resolution camera trap dataset from Central Europe. Over a 12-month period, 110 cameras were deployed across a forested area of 94 km 2 . Animal-to-camera distances were estimated via monocular depth estimation. To reduce manual annotation effort, we introduce an automated masking approach for calibration images based on the Segformer model, saving an estimated minimum of two minutes per image. Using red deer ( Cervus elaphus ) as model species, a full-factorial sensitivity study (n = 135 parameter configurations) was conducted, varying MegaDetector confidence threshold, depth sampling method, mean trigger intervals and right-hand distance truncation schemes. Results show that population density estimates are highly sensitive to the choice of depth sampling method, MegaDetector confidence threshold and mean trigger interval. Parameter-induced variation led to differences of up to 200% in population estimates relative to the lowest estimate, twice as much as the variation observed between spatial subsets of the study area differing in vegetation characteristics. However, relative differences between the two sub-areas remained largely consistent across configurations. These results highlight the critical impact of technical choices within CTDS workflows and emphasize the need for transparent reporting and standardization of key pipeline parameters. All CTDS analyses using automated pipelines should therefore be accompanied by dataset-specific sensitivity analyses to account for parameter uncertainty and validation challenges. Overall, our study demonstrates both the methodological challenges and practical applicability of automated CTDS for large-scale wildlife monitoring. • Automated masking reduces manual effort in depth calibration by ∼ 2 min per image. • CTDS parameter variation caused 200% difference, twice our spatial density variation. • Depth sampling, confidence thresholds and trigger intervals affect density estimates. • Automated CTDS requires standardization and dataset-specific sensitivity analyses.

Scientific research underpins effective conservation policy, yet current approaches for assessing whether scientific outputs meaningfully support defined management objectives rely primarily on manual expert review. This limitation constrains scalability, is time intensive and introduces potential bias in identifying knowledge gaps. We present a framework combining AI-assisted multi-target classification with systematic coverage analysis for automated evaluation of research alignment with conservation objectives. We compare traditional machine learning (TF-IDF + logistic regression), a generic BERT baseline, and an enhanced SciBERT approach incorporating domain-specific adaptations including multi-target architecture, balanced loss functions, and target weighting optimized for conservation science. The framework classifies research topics and conservation objective alignment, two dimensions requiring comprehension of scientific content and policy implications. We demonstrate the approach using 295 expert-annotated peer-reviewed studies from the Ross Sea region Marine Protected Area in Antarctica. Our enhanced multi-target SciBERT model achieved 70.0% macro F1, outperforming TF-IDF (59.5%) and BERT (52.0%) baselines, with per-target improvements of 21% on research topics and 14.5% on conservation objectives. The framework achieved 78% agreement with expert annotations, with particularly strong performance on conservation objective alignment (87.7% F1, 94% agreement). The integrated system successfully identified and quantified descriptive patterns in research coverage across thematic and policy dimensions, enabling systematic assessment for research prioritization and automated coverage analysis. While demonstrated in the Antarctic context, the framework architecture is broadly transferable, though successful adaptation requires retraining with domain-specific expert annotations and fine-tuning to match local management frameworks. • Enhanced multi-target SciBERT achieves 78% agreement with expert annotations for automated research-policy alignment assessment. • Framework reveals research coverage patterns with 59% of conservation topics appearing in <15 papers. • Multi-target architecture simultaneously classifies research themes and conservation objectives with 17.6% improvement over TF-IDF baseline. • Demonstrates scalable deployment across diverse contexts from Marine Protected Areas to SDG assessment. • Open-source implementation enables immediate adoption for evidence-based conservation planning.