Addax-Data-Science/NZI-ADS-v2

This repository contains open-source models redistributed for easy integration with AddaxAI, hosted by Addax Data Science. Each model retains its original license (see license files) and attribution. We comply with all original license terms. Users must review and comply with individual model licenses before use. See below for detailed model information including original sources, licenses, and attributions.

Owner

New Zealand Department of Conservation

Developer

Addax Data Science

Links

• Learn more
• License

About this model

New Zealand invasives b, the second New Zealand model in the AddaxAI model zoo. It identifies 24 species or higher-level taxons and was developed with the Department of Conservation, National Eradication Team, project reference 2025-12-NZI.

The first model, NZI-ADS-v1, stays available as New Zealand invasives a. Use a where the detection of rodents is critical, for example island biosecurity. Use b for broader coverage on everything else.

Training

• 2,117,539 camera trap images from 105 locations, all from local field deployments
• Split by camera location, not by image: 74 training, 21 validation, 10 test locations
• Largest classes capped at 300,000 images, selected per location with DINOv2 embeddings and farthest-point sampling rather than at random
• SpeciesNet EfficientNet V2 M backbone, fine-tuned, 52.8M parameters, 480x480 input
• On the held-out test set of 196,474 images: 94.1% accuracy, 94.5% precision, 94.1% recall
• The released checkpoint was retrained on the training, test and most of the validation data, so it has no held-out set of its own

The false detection class

One of the 24 classes is false detection, covering bait, stones, branches and moving vegetation that the detector reports as an animal. AddaxAI treats it as a non-label class, so those detections are dropped and never shown as a species. This keeps empty bait-station images out of the results. It also costs some rodent recall, because small rodents on bait stations at night look much like the things this class was trained on.

Files

• final-20260606.pt the fine-tuned checkpoint, carrying its own class names, input size, layout and normalisation
• always_crop_99710272_22x8_v12_epoch_00148.pt the SpeciesNet backbone. It supplies the module graph, which the fine-tuned state dict then fills in, so both files are required
• inference.py the shared addax-sppnet implementation, identical to the one in HWI-ADS-v1 and IND-ADS-v1
• taxonomy.csv taxonomic ranks per class, for the AddaxAI filter tree
• taxon-mapping.csv the training pipeline's class-to-GBIF mapping, which taxonomy.csv was resolved from

Nothing was stripped from either checkpoint. final-20260606.pt carries no optimizer or scheduler state.