[Paper Review] Computer vision-based estimation of invertebrate biomass

Mikko Impiö, Philipp M. Rehsen+6 more|arXiv (Cornell University)|Mar 6, 2026

Cell Image Analysis Techniques0 citations

TL;DR

The paper develops and evaluates two pathways for estimating dry mass of invertebrates from images: (i) a linear model using novel imaging-derived predictors (sinking speed and area) and (ii) end-to-end CNN models (single-view, multi-view, and metadata-aware) trained on BIODISCOVER image sequences, showing competitive, group-level biomass estimates without destructive weighing.

ABSTRACT

The ability to estimate invertebrate biomass using only images could help scaling up quantitative biodiversity monitoring efforts. Computer vision-based methods have the potential to omit the manual, time-consuming, and destructive process of dry weighing specimens. We present two approaches for dry mass estimation that do not require additional manual effort apart from imaging the specimens: fitting a linear model with novel predictors, automatically calculated by an imaging device, and training a family of end-to-end deep neural networks for the task, using single-view, multi-view, and metadata-aware architectures. We propose using area and sinking speed as predictors. These can be calculated with BIODISCOVER, which is a dual-camera system that captures image sequences of specimens sinking in an ethanol column. For this study, we collected a large dataset of dry mass measurement and image sequence pairs to train and evaluate models. We show that our methods can estimate specimen dry mass even with complex and visually diverse specimen morphologies. Combined with automatic taxonomic classification, our approach is an accurate method for group-level dry mass estimation, with a median percentage error of 10-20% for individuals. We highlight the importance of choosing appropriate evaluation metrics, and encourage using both percentage errors and absolute errors as metrics, because they measure different properties. We also explore different optimization losses, data augmentation methods, and model architectures for training deep-learning models.

Motivation & Objective

• Motivate scalable, non-destructive biomass estimation for invertebrates to support biodiversity monitoring.
• Introduce imaging-derived predictors (area and sinking speed) from BIODISCOVER as inputs for biomass modeling.
• Develop and compare linear models and CNN-based approaches (single-view, multi-view, metadata-aware) for dry mass estimation.
• Assess performance across taxonomic groups and in out-of-distribution scenarios.
• Propose evaluation metrics and reporting guidelines to enable fair model comparison.

Proposed method

• Use BIODISCOVER image sequences to compute predictors such as specimen area and sinking speed (derived from frame counts and positional data).
• Fit ordinary least squares linear models using (a) area only and (b) area plus sinking speed as predictors.
• Train end-to-end CNNs (single-view ResNet18, multi-view with two inputs, and metadata-aware variants) to map images (and optional metadata) to dry mass.
• Experiment with different loss functions (L1, L2, and percentage errors in log-space), augmentations, and model architectures (ResNet18 and EfficientNet variants).
• Evaluate models using multiple metrics (MAPE, MdAPE, MAE, RMSE, R^2 on log-transformed targets) and bootstrap confidence intervals.

(a) Araneae (O)

Experimental results

Research questions

• RQ1 Can image-derived features from BIODISCOVER predict dry mass accurately across diverse invertebrate morphologies?
• RQ2 Do CNN-based models outperform simple linear models using imaging predictors for biomass estimation?
• RQ3 How do multi-view and metadata-aware architectures compare to single-view models in biomass prediction?
• RQ4 What is the model behavior when generalizing to out-of-distribution taxa?
• RQ5 Is it beneficial to report both percentage-based and absolute error metrics for biomass estimation?

Key findings

• A linear model using area plus sinking speed (from image sequences) substantially improves RMSE and MAE over area-only models.
• Best linear model (area+speed) achieved RMSE 0.950 ± 0.20 and MAE 0.222 ± 0.03, with MdAPE 0.263 ± 0.01 and R2 0.921 ± 0.01.
• Among loss-function experiments, a percentual error objective yielded strong performance (MAPE 0.331 ± 0.01, MdAPE 0.267 ± 0.01, R2 0.906 ± 0.01).
• CNN-based models further improved performance, with metadata-aware ResNet18 trained on log-L1 loss (flips+90°) performing best in the optimization group.
• Ground-truth evaluation on larger heterogeneous Order dataset and smaller Species datasets showed competitive performance and demonstrated potential for end-to-end taxonomy-aware biomass estimation.

(b) Coleoptera (O)