Computer vision-based estimation of invertebrate biomass

Mikko Impiö; Philipp M. Rehsen; Jarrett Blair; Cecilie Mielec; Arne J. Beermann; Florian Leese; Toke T. Høye; Jenni Raitoharju

cs.CVMar 6, 2026

Computer vision-based estimation of invertebrate biomass

Authors: Mikko Impiö, Philipp M. Rehsen, Jarrett Blair, Cecilie Mielec, Arne J. Beermann, Florian Leese, Toke T. Høye, Jenni Raitoharju

Organizations: Finnish Environment Institute (Syke), Helsinki, Finland · University of Duisburg-Essen, Aquatic Ecosystem Research, Essen, Germany · University of Duisburg-Essen, Centre for Water and Environmental Research (ZWU), Essen, Germany · Aarhus University, Department of Ecoscience, Aarhus, Denmark, Aarhus, Denmark ·

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.

Figures & tables

Panel of Figure 1 — Figure 1: Randomly sampled images from both datasets grouped by taxonomic order. There is large variation in size and biomass within groups. S=Species dataset, O=Order dataset. The Species dataset was imaged completely separately, using a different device from the Order dataset, thus providing a test dataset for out-of-distribution generalization.

Figure 2: Overview of the BIODISCOVER imaging device setup. The device images specimens from two angles as they fall through an ethanol-filled cuvette. Useful metadata, such as sinking speed, can be calculated from the image timestamps, locations, and total number of images, as the camera frame rate is fixed.

Taxon	# spec.	# img.	m (mean)	m (median)	m (max)	m (min)	m (25p)	m (75p)	m (std)
Order dataset
Araneae	208	19 510	363.0	159.0	7 067.0	4.0	67.0	320.0	607.8
Coleoptera	204	17 440	673.5	213.0	26 013.0	19.0	67.0	479.0	2 025.1
Diptera	192	27 055	165.0	58.0	1 675.0	4.0	24.0	164.0	259.7
Hemiptera	175	17 915	365.3	80.0	20 695.0	9.0	27.0	345.0	1 045.3
Hymenoptera	201	26 852	211.3	93.0	3 474.0	2.0	30.0	175.0	406.5

Table 1: Overview of our datasets. Dry masses (m) are presented in micrograms (µg). 25p=25th percentile, 75p=75th percentile, std=Standard deviation.

Dataset	# of specimens	# of weighed specimens	% weighed
Ärje et al. (2020)	533	65	12.2
Schneider et al. (2022)	13 059	68	0.5
Rehsen et al. (2026)	743	743	100.0
This study	1 116	1 116	100.0

Table 2: Comparison of our dataset with previous similar invertebrate biomass datasets in the literature.

Panel of Figure 3 — Figure 3: Histogram of dry mass measurements (a), and relationships between mass measurements and dataset features (b-d).

Figure 4: Area and dry mass ranges of Asellus aquaticus specimens. The x-axis shows samples from five mass quantiles (0-20%, 20%-40%, 40%-60%, 60%-80%, and 80%-100%) and the y-axis shows minimum, mean and maximum area from these quantiles. Below each image, the specimen mass (m) and area (A) of the specimen in this particular image is given.

Figure 5: Overview of the different CNN models used in this paper. Image inputs

x

are passed to CNN encoders (

g

), while possible metadata

v

is passed to a fully connected metadata encoder

\mu

. Encoders output intermediate feature vectors

z

. For the multi-view and metadata-aware architectures, feature vectors are concatenated before passing the features to a fully connected projection head (

h

) with either one or two layers. The output of this network is the final biomass estimate

\hat{y}

Linear models
	Absolute		Percentual
Name	RMSE	MAE	MAPE	MdAPE	R2
Linear (area)	1.094 ± 0.21	0.289 ± 0.03	0.578 ± 0.02	0.394 ± 0.01	0.820 ± 0.01
Linear (area + speed)	0.950 ± 0.20	0.222 ± 0.03	0.341 ± 0.01	0.263 ± 0.01	0.921 ± 0.01
Loss functions
Percentual error	0.871 ± 0.14	0.228 ± 0.03	0.331 ± 0.01	0.267 ± 0.01	0.906 ± 0.01

Table 3: Full results of model optimization experiments. The overall best performing model was the metadata-aware ResNet18 trained with log-L1 loss, flips, and 90° rotations. Best performance of each group is underlined for each metric. Best overall performance for each metric is marked with bold . Bootstrapped standard deviations are reported.

Panel of Figure 6 — Figure 6: Comparison of all trained models with bootstrapped 95% confidence intervals.

In-distribution evaluation
	Order Dataset		Species Dataset
			Asellus aquaticus		Caenis horaria		Kageronia fuscogrisea
Method	MAE	MdAPE	MAE	MdAPE	MAE	MdAPE	MAE	MdAPE
Lin. A (Baseline)	0.289 ± 0.03	0.394 ± 0.01	0.144 ± 0.02	0.116 ± 0.03	0.035 ± 0.01	0.403 ± 0.06	0.070 ± 0.01	0.209 ± 0.04
Lin. A+S	0.222 ± 0.03	0.263 ± 0.01	0.126 ± 0.02	0.126 ± 0.02	0.022 ± 0.00	0.234 ± 0.03	0.079 ± 0.02	0.226 ± 0.04
CNN	0.173 ± 0.02	0.217 ± 0.01	0.198 ± 0.03	0.202 ± 0.04	0.038 ± 0.01	0.324 ± 0.05	0.082 ± 0.01	0.302 ± 0.05

Table 4: Biomass estimation results. Lin. A = Linear model using only area as predictor. Lin. A+S = Linear model with area and speed predictors. CNN = CNN model trained with in-domain data. CNN A+S = Metadata-aware CNN with area+speed predictors. FT CNN = CNN model pretrained with OOD data and fine-tuned with in-domain data. FT CNN A+S = Similarly fine-tuned metadata-aware CNN model. FT CNN Frz = FT model with convolutional layers frozen during training. FT CNN A+S Frz = Metadata-aware model with CNN encoder and metadata encoder frozen. OOD = CNN model trained with out-of-domain data. OOD A+S = Metadata-aware CNN model trained with out-of-domain data. Lin. OOD = Linear A+S model trained with OOD data. Best models for each group are marked with bold . Bootstrapped standard deviations are reported.

Figure 7: Distribution of log-transformed dry mass predictions against true values, with residuals calculated on the log-transformed scale. Pearson’s correlation coefficients (r) are shown for each method. 1:1 line is shown for reference. All models overestimate small samples and underestimate larger samples.

Outlier taxon	Model	RMSE	MAE	MAPE	MdAPE	R2
Araneae	Linear	0.262 ± 0.05	0.120 ± 0.02	0.265 ± 0.02	0.191 ± 0.02	0.938 ± 0.01
Araneae	CNN	0.450 ± 0.08	0.212 ± 0.03	0.491 ± 0.05	0.342 ± 0.04	0.840 ± 0.02
	CNN+Meta	0.246 ± 0.05	0.121 ± 0.02	0.264 ± 0.02	0.204 ± 0.01	0.930 ± 0.01
Coleoptera	Linear	1.873 ± 0.52	0.508 ± 0.14	0.345 ± 0.02	0.304 ± 0.02	0.903 ± 0.01
Coleoptera	CNN	1.683 ± 0.39	0.545 ± 0.12	0.391 ± 0.03	0.316 ± 0.03	0.882 ± 0.02
	CNN+Meta	2.333 ± 0.60	0.657 ± 0.17	0.397 ± 0.02	0.359 ± 0.04	0.868 ± 0.01

Table 5: Generalization performance on the order dataset. The models are trained on all other classes except for the one in the Outlier taxon -column, and tested on the Outlier taxon only. Bootstrapped standard deviations are reported.

Figure 8: Classification confusion matrix, as percentages of the true class. See Table 1 for number of specimens for each class. Ara=Araneae, Col=Coleoptera, Dip=Diptera, Hem=Hemiptera, Hym=Hymenoptera

Figure 9: Swarm plot of classification combined with biomass estimation. Each specimen is represented as a dot. True dry masses are plotted next to predicted values. Specimen dry masses are plotted based on their predicted class. Classification mistakes are highlighted with a darker color. Each dot corresponds to a single specimen. Two-sample Kolmogorov-Smirnov test statistics are reported for each group.

	Precision	Recall	F1-score
Araneae	0.986	0.990	0.988
Coleoptera	0.990	0.995	0.993
Diptera	0.968	0.938	0.952
Hemiptera	0.982	0.943	0.962
Hymenoptera	0.915	0.965	0.939

Table 6: Classification metrics for the Order dataset.

Biodiversity Monitoring

See all 23 papers

Deep learning-based hierarchical insect classification using camera trap imagery
Jul 30, 2026Zaki Mahfoud, Juan A. Chiavassa, Simon Walther +2Hierarchical Classification Wildlife Monitoring
Weihenstephan-Triesdorf University of Applied Sciences, Department of Sustainable Agriculture and Energy Systems, Professorship of Smart Farming, Am Staudengarten 1, 85354 Freising, Germany · Weihenstephan-Triesdorf University of Applied Sciences, Institute for Biomass Research, Neuseser Str. 1, 91732 Merkendorf, Germany · Weihenstephan-Triesdorf University of Applied Sciences, Department of Sustainable Agriculture and Energy Systems, Professorship of Digital Farm Management, Am Staudengarten 1, 85354 Freising, Germany
Automated high-frequency quantification of fish communities and biomass using computer vision
May 11, 2026Kota Ishikawa, Takuma Masui, Keita Koeda +3Wildlife Monitoring
Graduate School of Life Sciences, Tohoku University, Sendai, Japan · Advanced Institute for Marine Ecosystem Change (WPI-AIMEC), Tohoku University, Sendai, Japan · Graduate School of Science and Engineering, University of the Ryukyus, Okinawa, Japan +1
WildProp: Visual Estimation of Wildlife Body Proportions at Scale
Jun 30, 2026Mustafa Chasmai, Aaron Sun, Subhransu MajiImage Retrieval
Manning College of Information and Computer Sciences University of Massachusetts, Amherst

Jul 30, 2026J/K move · Enter open · S save

Deep learning-based hierarchical insect classification using camera trap imagery

Zaki Mahfoud, Juan A. Chiavassa, Simon Walther, Florian Haselbeck+1

Weihenstephan-Triesdorf University of Applied Sciences, Department of Sustainable Agriculture and Energy Systems, Professorship of Smart Farming, Am Staudengarten 1, 85354 Freising, Germany · Weihenstephan-Triesdorf University of Applied Sciences, Institute for Biomass Research, Neuseser Str. 1, 91732 Merkendorf, Germany · Weihenstephan-Triesdorf University of Applied Sciences, Department of Sustainable Agriculture and Energy Systems, Professorship of Digital Farm Management, Am Staudengarten 1, 85354 Freising, Germany

Hierarchical Classification Wildlife Monitoring Biodiversity Monitoring

Open paper AskPDF arXiv

Declining insect populations make reliable biodiversity monitoring increasingly urgent, yet monitoring of insect biodiversity is hampered by a lack of standardised data and by costly and time-consuming manual identification by expert entomologists. Deep learning-based image classifiers, processing data from automated non-lethal camera traps, have the potential to transform and scale insect biodiversity monitoring. However, challenges remain in acquiring expert-annotated datasets, developing model architectures that generalise well across diverse taxonomic levels and training models on highly imbalanced data. Hierarchical data also benefits from designing models that default to higher-confidence, coarser-level predictions, when uncertain about finer taxonomic levels. In this paper we address these challenges with a deep learning-based hierarchical classification model. First, we present a manually curated, long-tailed dataset of around one million images of insects, extracted from 1,801 camera-trap video recordings and annotated with a five-level, 34-class hierarchy. Further, we adapt a hierarchical classification model architecture to a five-level variable-depth hierarchy, with class-balanced weighting. Our model improves on non-hierarchical classifiers by leveraging biological taxonomy to extract granularity-specific visual features and makes hierarchy-consistent predictions to the deepest taxonomic level that meets a confidence threshold (T = 0.6). Our model achieved a per-level accuracy of 80-99% on test data, across five levels of hierarchy. Furthermore ...

Abstract

Biodiversity Monitoring

Explore similar work

Deep learning-based hierarchical insect classification using camera trap imagery

Automated high-frequency quantification of fish communities and biomass using computer vision

WildProp: Visual Estimation of Wildlife Body Proportions at Scale