BEEomass

EfficientNetV2 regression model that estimates the dry biomass of individual arthropods from images.

Intended use. Estimating individual dry biomass of terrestrial arthropods from segmented images of single specimens, as an alternative to drying and weighing each one. It is designed for images from EntoScan, a modified flatbed scanner, but the authors report that it generalises to other calibrated imaging setups (it was also trained and tested on Biodiscover images). The model does not predict mass directly. It predicts a size-normalised Biomass Factor (BF), and dry mass is recovered from BF using the image’s known physical scale.

Architecture and training

• Architecture: EfficientNetV2 (variant v2_s), ImageNet-pretrained, with the final layer replaced by a single-output regression head (dropout 0.6 → linear → scalar)
• Task: Regression (one Biomass Factor value per image)
• Target: BF = M^(1/3) / L, in mg^(1/3) mm^-1, where M is dry mass (mg) and L is the physical width of the 224 px field of view, L = 25.4 × (224 / s) / DPI (mm), with s the resize factor applied to the segmented image. Under isometric scaling BF is constant, so it captures body “compactness” rather than absolute size
• Trained on: paired images and dry biomass measurements from EntoScan (field-collected specimens from pan traps in Denmark, Malaise traps in Sweden, and vacuum sampling, plus lab-reared Drosophila) and the Biodiscover dataset (2,037 images). Each specimen was imaged several times in different poses. Data were split 80/10/10 into train, validation and test sets at the specimen level. The combined training data holds 9,481 images of 3,142 weighed specimens
• Taxonomic coverage: terrestrial arthropods, mainly insects (e.g. Diptera, Hymenoptera, Coleoptera, Plecoptera, Ephemeroptera, Trichoptera) plus spiders. The model is taxon-agnostic and does not classify specimens
• Loss and optimisation: MSE on BF. AdamW (learning rate 1e-4, weight decay 1e-5), batch size 32, 500 epochs, StepLR halving the learning rate every 100 epochs. The checkpoint with the lowest validation loss was kept
• Augmentation (training only): scale augmentation (image shrunk by s ~ U(0.5, 1) onto a white 224 × 224 canvas, with the BF target adjusted to match), random horizontal flips, 90° rotations, colour jitter, a random choice of Gaussian blur or sharpening, and a ±20% multiplicative jitter on the target
• Framework: PyTorch

Inputs and outputs

• Input size: 224x224. Specimens are first segmented from the background (the authors use FlatBug), resized so the longest side is 224 px with aspect ratio preserved, then padded with white to a square
• Channel order: RGB
• Outputs: one scalar Biomass Factor per image. Dry mass is recovered as M = (BF × L)^3, which requires the acquisition resolution (DPI) and resize factor for each image
• Input modality: image only (discriminative model)

Performance

Values from the source publication, measured on the held-out test sets. Confidence intervals (95%) come from 500 bootstrap iterations, sampling one image per specimen. The baseline is a linear model on the square root of segmented area.

The authors also demonstrate the model in two case studies: detecting temperature-induced size differences in five lab-reared Drosophila species, and tracking seasonal biomass dynamics in field-collected Pachygnatha degeeri spiders.

Limitations

• The Biomass Factor assumes a fairly consistent size–mass relationship. Taxa that depart strongly from isometric scaling, or have unusual body forms, may be predicted less accurately. Taxa outside the training data have not been evaluated
• Predictions depend on accurate segmentation. Segmentation errors, especially in dense or overlapping bulk samples, carry through to the biomass estimate
• Dry mass can only be recovered when the physical scale (DPI) of the image is known, so images need a calibrated setup
• The source publication is a preprint and has not been peer reviewed

License and rights

• License: CC-BY-4.0
• Commercial use: allowed
• Restrictions: Attribution is required by the license. Cite the source publication listed under Citation below.
• Upstream source: https://zenodo.org/records/22097250
• Code: https://github.com/darsa-group/BEEomass

Citation

@article{baghooee2026beeomass,
  title   = {EntoScan and BEEomass: a standardized imaging system and a physically
             motivated model for high-throughput dry biomass estimation of arthropods},
  author  = {Baghooee, Melika and Thalheim, Robert and Hasan, Fevziye and Toft, S{\o}ren
             and Kristensen, Torsten Nyg{\aa}rd and Geissmann, Quentin},
  journal = {EcoEvoRxiv},
  year    = {2026},
  doi     = {10.32942/X2Q687},
  note    = {Preprint}
}

How to obtain the weights

From Zenodo record 22097250. No authentication is required. The weights are in the 2026-08-25_BEEomass_model.zip file. Training and inference scripts (02-train.py, 03-predict.py) are in the BEEomass GitHub repository.