AMI moth classifiers

Names moths to species from light-trap crops, with a regional model per checklist, plus a moth/non-moth filter and an order-level classifier for everything else.

Intended use. The classification stages of the AMI (Automated Monitoring of Insects) pipeline, run by the AMI data companion and the Antenna platform. Each crop from the AMI insect detector is processed in three steps:

1. The moth/non-moth filter checks it.
2. Moths go to a species classifier: pick the regional model matching your checklist, or the global one.
3. Optionally, the order classifier labels the other arthropods.

Architecture and training

• Species classifiers: ResNet50 (timm, ImageNet-initialised), trained on AMI-GBIF, which is GBIF occurrence images for regional moth checklists.
  • Settings: 128 px, 30 epochs, RandAugment plus MixRes (mixed-resolution augmentation, which simulates low-resolution trap crops).
  • The April 2024 NE-America, W-Europe and Panama files are the ResNet50 baselines of Jain et al. (ECCV 2024); their species counts match the paper’s Table 1.
  • The paper’s best model (ConvNeXt-B) is not the one deployed.
  • The global model (29,176 species, August 2024) follows the same recipe.
• Turing models (Costa Rica, Kenya-Uganda, Anguilla): ResNet50 at 300 px, trained by the Alan Turing Institute team with the AMI tooling.
• Moth/non-moth: ResNet50 at 128 px, trained on AMI-GBIF moths plus about 350k non-moth images.
• Order classifier: ConvNeXt-T (convnext_tiny_in22k) at 128 px, with 16 arthropod orders, trained at Mila in January 2025.
• Training code: RolnickLab/ami-ml.

Inputs and outputs

• Input: one crop per call, square-resized to the model’s input size (128 or 300 px), with ImageNet normalisation.
• Output: softmax scores over the model’s label map.
• Label maps are separate JSON files at the same base URL (named in each asset note) and map indices to species names or GBIF keys.
• In the AMI API each classification carries the top label, the full labels, scores and logits vectors, and a taxon_rank: SPECIES, SUPERFAMILY for the binary model, or ORDER.

Performance

Jain et al. (ECCV 2024) report micro top-1 accuracy (in %) on AMI-Traps: 2,893 expert-annotated trap images and 52,948 insects from NE America, W Europe and Central America. The model is the deployed ResNet50 architecture.

Moth/non-moth ResNet50 on all 51,120 AMI-Traps crops: accuracy 86.48, precision 68.31, recall 95.03, F1 79.48. On crops larger than 150 px: accuracy 92.95, F1 93.70.

The global model, the Turing models, the order classifier and the Panama-2023 model have no published metrics.

Limitations

• Closed set, Lepidoptera only (species models). A moth not on the regional checklist still receives one of its species. Match the model to your region.
• Domain gap. Trained on human-taken GBIF photos. Accuracy on held-out GBIF images is about 87 %, against 72–79 % on real trap crops.
• Long tail. Species with few training images reach only 27–49 % on AMI-GBIF.
• Low validation coverage. The Central America and Turing models have little or no in-the-wild validation.
• Small crops. Moth/non-moth precision drops on small crops, most of which are non-moths.
• Not the best model. The deployed models are ResNet50 baselines, not the paper’s best-performing ConvNeXt-B.
• Hosting. The weights live only on a Digital Research Alliance of Canada object store and are unpinned. A legacy file (Turing-UK) already returns 404.

How to obtain the weights

Direct HTTPS download from the object store, with no authentication. The AUTH_… segment of the URL is required. The paths are defined in trapdata/ml/models/classification.py and the base URL in trapdata/settings.py of ami-data-companion, curated at commit 55d0787570f8c9d92c450f4fb65a58dd25cf6634. trapdata downloads them automatically. Each classifier needs its label-map JSON, from the same base URL.

The Alan Turing Institute models (Costa Rica, 5,952 species; Kenya-Uganda, 3,045; Anguilla, 79; ResNet50 at 300 px, no published metrics) are at the same base URL: moths/classification/turing-{costarica_v03_resnet50_2024-06-04-16-17,kenya-uganda_v01_resnet50_2024-11-19-18-44,anguilla_v01_resnet50_2024-06-28-17-01}_state.pt.

Older models are defined in the same file but not exposed by the API:

• UK-Denmark mixres 2023
• Quebec-Vermont mixres 2022
• Panama 2022 (148 species)
• Panama 2023 (1,060 species)
• the 2022 EfficientNetV2-B3 moth/non-moth model

License: no separate license is stated for the weights. This entry records AGPL-3.0, the license of the AMI code that defines, trains and serves them (ami-data-companion, ami-ml, antenna). The AMI dataset is MIT (Zenodo 10.5281/zenodo.11358689).

Citation

@inproceedings{jain2024ami,
  title     = {Insect Identification in the Wild: The {AMI} Dataset},
  author    = {Jain, Aditya and Cunha, Fagner and Bunsen, Michael J. and Ca{\~n}as, Juan Sebasti{\'a}n and
               Pasi, L{\'e}onard and Pinoy, Nathan and Helsing, Flemming and Russo, JoAnne and
               Botham, Marc S. and Sabourin, Michael and Fr{\'e}chette, Jonathan and Anctil, Alexandre and
               Lopez, Yacksecari and Navarro, Eduardo and P{\'e}rez, Filonila and Zamora, Ana C. and
               Ramirez-Silva, Jose Alejandro and Gagnon, Jonathan and August, Tom A. and Bjerge, Kim and
               Gomez Segura, Alba and B{\'e}lisle, Marc and Basset, Yves and McFarland, Kent P. and
               Roy, David B. and H{\o}ye, Toke T. and Larriv{\'e}e, Maxim and Rolnick, David},
  booktitle = {Computer Vision -- ECCV 2024},
  series    = {Lecture Notes in Computer Science},
  volume    = {15095},
  pages     = {55--73},
  publisher = {Springer, Cham},
  year      = {2024},
  doi       = {10.1007/978-3-031-72913-3_4}
}