DeepForestVisionV2: Ecology-Driven Taxonomy Expansion for Camera-Trap Monitoring in African Tropical Forests

Technical Summary: DeepForestVisionV2

Problem Statement
Camera-trap monitoring in African tropical forests is increasingly expanding beyond closed-canopy interiors to include riverbanks, clearings, and park edges. While automated species identification is essential for processing the growing volume of data, existing tools often lack the taxonomic resolution required for these diverse deployment scenarios. The original DeepForestVision model, while superior to other open tools (Mbaza, Zamba, SpeciesNet) in offline video and photo processing, was designed for closed-canopy, ground-level interiors. Its 35-class prediction space is too coarse for emerging ecological gradients:

1. Vertical Stratification: Arboreal primates become visible near clearings and canopy openings.
2. Scene Openness: Riverbanks expose birds and semi-aquatic taxa rarely seen in dense interiors.
3. Anthropogenic Interfaces: Park-edge monitoring requires distinguishing wildlife from livestock to avoid nuisance alarms.

The original model's coarse labels (e.g., generic "bird," "monkey," or "civet/genet") are insufficient for ecological interpretation or management actions in these specific contexts.

Methodology
The authors present DeepForestVisionV2, an ecology-driven expansion of the original system that retains the same offline, no-code workflow (via AddaxAI) for both photographs and videos.

• Taxonomy Expansion: The prediction space was expanded from 35 to 64 classes (61 animal classes, plus human, vehicle, and blank). This redesign specifically targets the three ecological gradients mentioned above, introducing finer-grained classes for primates (e.g., specific mangabey and colobus species), birds (e.g., cranes, hornbills), semi-riparian taxa (e.g., otters, hippos), and domestic animals (cattle, dogs, goats).
• Datasets: The model was trained on a multi-country corpus comprising 1,535,010 photographs and 243,354 videos from African tropical forest projects across Uganda, Gabon, Sierra Leone, and the Congo. Data sources include the Sebitoli Chimpanzee Project, CIRAD, Biotope, and Wildlife Insights.
• Pipeline Architecture:
  • Detection: MegaDetector v5 is used to detect animals, humans, and vehicles. Human and vehicle labels are assigned directly; animal detections are cropped.
  • Classification: Cropped animal detections are processed by a fine-tuned DINOv3 ViT-B/16 classifier.
  • Aggregation: For videos, class scores are computed on sampled frames and averaged to produce a single item-level prediction.
• Training Strategy: The classifier was fine-tuned using PyTorch with extensive augmentation (random resized crops, flips, color jitter, grayscale, etc.). Optimization used AdamW with separate learning rates for the backbone and head, label smoothing, and cosine scheduling.
• Evaluation Protocol:
  • Validation: A cross-country cropped-photo validation set (derived from training sources but split at the camera level) was used to assess robustness across sites and settings.
  • Deployment Benchmarks: Three held-out video benchmarks from Kibale National Park, Uganda, were used to simulate real-world conditions: Forest Interior (17,899 videos), Riverbanks (1,016 videos), and Park Edge (146 videos).
  • Metrics: Accuracy, balanced accuracy, macro-F1, and per-class precision/recall were reported. For the park-edge benchmark, an "alarm task" was defined to measure the system's ability to distinguish primate intrusions from nuisance goat passages.

Key Contributions

1. Ecology-Driven Taxonomy: An expansion from 35 to 64 classes aligned with vertical stratification, scene openness, and anthropogenic interfaces.
2. Robustness Benchmark: A geographically diverse, Africa-wide cropped-photo validation set to assess classifier robustness across different sites and camera-trap settings.
3. Deployment-Centric Evaluation: A comparative evaluation of DeepForestVisionV2 against the original DeepForestVision on three specific video benchmarks, demonstrating performance under matched offline workflows.

Results

• Validation Set: DeepForestVisionV2 achieved 0.86 accuracy, 0.81 balanced accuracy, and 0.82 macro-F1 across the full 61-animal target space, demonstrating robustness across countries and habitats.
• Forest Interior Benchmark: The model maintained the original accuracy (0.89) while increasing the number of identified taxa from 22 to 29, surfacing specific arboreal primates and bird groups previously merged into coarse labels.
• Riverbank Benchmark: Accuracy remained comparable (0.72 vs. 0.73), but the number of identified taxa increased from 4 to 9, capturing birds and semi-aquatic species.
• Park-Edge Benchmark: The model significantly improved accuracy from 0.62 to 0.86. Crucially, it reduced false alarms (goat detections triggering primate alerts) from 11 to 0, while maintaining high recall (0.95) for true primate alarms.
• Calibration: The model exhibited an Expected Calibration Error (ECE) of 0.17, indicating under-confidence (empirical accuracy exceeded predicted confidence), suggesting that confidence scores should not be treated as calibrated probabilities without post-hoc adjustment.

Significance and Claims
The paper claims that DeepForestVisionV2 materially improves the field utility of camera-trap monitoring without sacrificing the robustness of the original offline workflow. By aligning the taxonomy with specific ecological deployment gradients, the system:

• Preserves or improves baseline accuracy despite the increased difficulty of a 64-class task.
• Provides necessary taxonomic resolution for downstream ecological questions in non-canopy environments.
• Offers direct operational benefits for park management by filtering nuisance alarms (livestock) at park edges, making the system more viable for automated alerting pipelines where false positives are costly.

The authors conclude that ecology-driven taxonomy expansion makes field-ready AI workflows more informative for conservation across contrasting habitats, though they note that performance on video benchmarks is lower than on cropped images due to detection errors and limited validation support for rare taxa.