BioMoth

Counts moths on blacklight-sheet photos and estimates their biomass from box sizes.

Intended use. Non-lethal monitoring of moth abundance and biomass from a camera photographing a UV-lit sheet at fixed intervals through the night. It was built to replace lethal bucket light traps in long-term ecological studies. It answers how many moths, and how much moth mass, not which moths.

Architecture and training

• Architecture: Ultralytics YOLO11s, one class (Moth), 640x640 input.
• Trained on: blacklight-sheet photos from Hubbard Brook Experimental Forest (NH, USA).
  • Five watershed sites, mid-June to mid-September 2023.
  • One photo every 10 minutes from 22:00 to 04:00, with a single Nikon camera setup.
• Data: the raw images are published on EDI (see Citation). The training/validation split and the box annotations are not released.
• Biomass step: this is not a learned model.
  • Box width and height are converted to cm² with a hard-coded camera-to-sheet geometry (130 cm distance, about 44x83 cm frame).
  • Each box goes into one of four log2 area bins, and each bin gets a fixed mean mass (2.75, 54, 361 or 418 mg). The masses are summed.

Inputs and outputs

• Input: RGB sheet photos, letterboxed to 640. The source photos are about 6000 px wide.
• Output: A custom batch CSV containing boxes, confidence, class and measurements after TorchScript inference and author postprocessing.
• Thresholds: the README recommends conf = 0.23, while the inference notebook uses 0.25, with NMS IoU 0.5.
• Counts and biomass are aggregated per image and per night in the runTrends notebook.

The documented BioMoth CSV profile follows runInference.ipynb at revision 1d631ce. Its active batch export writes 15 columns, including original-image pixel boxes, confidence, class, image/site metadata and calibrated dimensions. A CPU probe of the real TorchScript checkpoint and original notebook functions produced 32 rows on a specimen fixture and none on a blank control. The hard-coded year and calibration factors are preserved; this does not validate physical measurements on that fixture.

Performance

It is not confirmed whether the repository checkpoint (dated June 2025) is the exact one behind the preprint’s numbers.

Limitations

• Single class. All Lepidoptera become Moth. There is no family, genus or species output, and no reported analysis of other insects (caddisflies, beetles) being counted as moths.
• Coarse biomass. Four fixed-mass bins from 2-D box area, with no per-species allometry. The pixel-to-cm factor assumes the authors’ exact camera distance and framing, so it must be recalibrated for any other setup.
• Small moths. Photos are about 6000 px wide and are downscaled to 640. Very small moths are likely to be missed. This is not a reported result.
• Relative index, not absolute counts. Detection is per frame with no tracking, so a resting moth is counted again in every 10-minute frame.
• One site network. One forest, one season, one camera setup. It has not been validated elsewhere.
• Preprint. The paper is not yet peer-reviewed.

How to obtain the weights

A plain file in the GitHub repository (not Git LFS), downloadable without login. The asset link is pinned to commit 1d631ce7af4c0fa8eae5d60d8f2ec352fa2644f3. It loads with ultralytics.YOLO("best.torchscript", task="detect").

License: the repository is licensed Apache-2.0. The weight file’s embedded Ultralytics metadata reads AGPL-3.0, as every Ultralytics YOLO11 export does. Commercial use is therefore recorded as unknown. Check with the authors, and with the Ultralytics license, before commercial use.

Citation

@article{lutz_biomoth_2026,
  author  = {Lutz, David A. and Suhavi, Suhavi and Jones, Jessica S. and Aldrich, Quinn M. and
             Cummings, Wyatt J. and Cui, Kangning and Ayres, Matthew P.},
  title   = {{BioMoth}: Non-Lethal Estimation of Moth Abundance and Biomass from Blacklight Sheet
             Imagery Using Computer Vision},
  journal = {SSRN preprint},
  year    = {2026},
  doi     = {10.2139/ssrn.6943698}
}

Images: Suhavi, S., D.A. Lutz, J.S. Jones, M.P. Ayres. 2026. Blacklight Sheet Invertebrate Photos at Hubbard Brook Experimental Forest; 2023 ver 1. Environmental Data Initiative. https://doi.org/10.6073/pasta/7ac5818bb45bb42c2d935ce7e3756c00