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

2026-06-18 · Source: Computer Vision and Pattern Recognition · Field: Science & Research — Life Sciences & Biology, Environmental Science & Earth Systems, Artificial Intelligence & Machine Learning · Depth: Advanced, quick

Summary

DeepForestVisionV2 is an ecology-driven expansion of the DeepForestVision camera-trap classification tool, increasing its prediction space from 35 to 64 classes (61 animal, plus human, vehicle, and blank). This update addresses limitations of the previous version, which was designed for closed-canopy interiors, by accommodating vertical stratification, scene openness, and anthropogenic interfaces encountered in diverse African tropical forest environments like riverbanks and park edges. DeepForestVisionV2 maintains the existing offline workflow and was trained on a large dataset of 1,535,010 photographs and 243,354 videos from multi-country projects. Evaluation on a cross-country validation set yielded 0.86 accuracy, 0.82 macro-F1, and 0.81 balanced accuracy. Furthermore, deployment benchmarks showed it preserved or improved baseline accuracy, increasing identified taxa from 22 to 29 in forest-interior videos and from 4 to 9 at riverbanks, while boosting park-edge accuracy from 0.62 to 0.86 and eliminating 11 false alarms.

Key takeaway

For ecologists managing camera-trap projects in African tropical forests, especially those expanding monitoring to riverbanks, clearings, or park edges, DeepForestVisionV2 provides a critical upgrade. You can now identify 64 distinct classes, including arboreal primates and human-associated confounders, with higher accuracy and fewer false alarms. This tool enhances your ability to monitor diverse ecological gradients effectively, preserving robustness across sites and camera settings.

Key insights

DeepForestVisionV2 expands camera-trap classification to 64 classes, improving utility across diverse African tropical forest habitats.

Principles

• Taxonomy expansion improves utility for diverse ecological gradients.
• Offline workflows are crucial for remote camera-trap deployments.
• Comprehensive training data enhances robustness across sites.

Method

DeepForestVisionV2 expands its prediction space from 35 to 64 classes, including 61 animal taxa, human, vehicle, and blank. It uses an offline workflow, trained on 1,535,010 photographs and 243,354 videos.

In practice

• Deploy camera traps in varied vertical and scene openness gradients.
• Incorporate human and livestock classes for park-edge monitoring.
• Utilize offline classification tools for remote African forest projects.

Topics

• Camera-Trap Monitoring
• DeepForestVisionV2
• Tropical Forest Ecology
• Wildlife Classification
• Computer Vision
• Biodiversity Monitoring

Best for: Computer Vision Engineer, AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Where wild things roam: Identifying wildlife with SpeciesNet — SpeciesNet is an open-source AI model for automated wildlife identification in camera trap images.
• We’re announcing Canopy Height Maps v2 (CHMv2), an open source model for high-resolution global forest canopy mapping, developed in partnership with the @WorldResources. CHMv2 leverages our DINOv3 Sat-L vision model, specifically optimized for s - x.com — CHMv2, an open-source model, uses DINOv3 Sat-L for high-resolution global forest canopy mapping.
• Visual Place Recognition in Forests with Depth-Aware Distillation — Depth-aware distillation enhances forest Visual Place Recognition by integrating geometric cues into DINOv2, improving robustness.
• Meta AI Releases Sapiens2: A High-Resolution Human-Centric Vision Model for Pose, Segmentation, Normals, Pointmap, and Albedo — Sapiens2 unifies multiple human vision tasks into a single model, improving performance and simplifying pipelines.
• USU-Corn-WeedDB: A UAV RGB Image Dataset for Multi-Species Weed Detection in Forage Corn — A new UAV RGB dataset enables multi-species weed detection in forage corn, supporting supervised and semi-supervised deep learning.
• SegmentAnyTreeV2: Scaling Transformer-Based Tree Instance Segmentation Across Sensors, Platforms, and Forests — SegmentAnyTreeV2 offers robust, scalable tree instance segmentation for forest point clouds using a Transformer-based architecture.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.