🌿 ecology

COLLEMBOT: AI-Based Counting of Collembola for OECD 232 Tests

The paper introduces COLLEMBOT, an AI-based automated counting tool using YOLOv11 that significantly reduces the time and labor required for OECD 232 soil toxicity tests while maintaining high accuracy and regulatory reliability compared to manual counting.

Original authors: Wehrli, M., Meyer, A. F., Souza da Silva, E., van Loon, S., van Hall, B. G., van Gestel, C. A. M., Natal-da-Luz, T., Doering, M. V. R., Feldhaar, H., Mair, M., Jordan, D., Langer, M.

Published 2026-01-20

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Wehrli, M., Meyer, A. F., Souza da Silva, E., van Loon, S., van Hall, B. G., van Gestel, C. A. M., Natal-da-Luz, T., Doering, M. V. R., Feldhaar, H., Mair, M., Jordan, D., Langer, M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a scientist trying to check if a new chemical is safe for the soil. To do this, you need to count thousands of tiny, spring-loaded bugs called Collembola (specifically Folsomia candida). These bugs are like the "canaries in the coal mine" for soil health; if they survive a chemical test, the soil is likely safe for other life.

For decades, scientists have had to do this counting by hand. It's like trying to count every single grain of sand on a beach while wearing thick gloves. It takes forever, it's exhausting, and because humans get tired, they sometimes make mistakes or count differently than the person next to them. This slow, error-prone process means we don't have enough data to make good decisions about chemical safety.

Enter COLLEMBOT.

Think of COLLEMBOT as a super-powered, tireless digital assistant that acts like a highly trained eagle eye. Instead of a human squinting at a petri dish, this tool uses a special type of artificial intelligence (called a YOLOv11 neural network) to "see" the bugs in photos.

Here is how it works in plain terms:

The Training: The AI didn't just guess. It was taught by looking at over 3,200 high-quality photos of these bugs taken in five different countries (from the Netherlands to Denmark). It learned to recognize the bugs whether they were in standard soil mixtures or different types of dirt.
The Proof: The scientists tested this digital assistant against real humans. The results were like matching two identical twins: the AI's counts matched the human counts almost perfectly (with a match score between 88% and 99%). When they used the AI's numbers to calculate if a chemical was harmful, the results were almost identical to the human calculations.
The Speed Boost: This is where the magic really happens. Imagine a task that takes a human team about 137 hours (that's nearly two full weeks of non-stop work) to finish. COLLEMBOT can do the exact same job in less than 3 hours. That is a 97% time savings. It's the difference between walking across a country and taking a bullet train.

The Bottom Line
COLLEMBOT doesn't change the rules of the game; it just plays the game much faster and more consistently. By automating the counting, it removes the human fatigue and bias, allowing scientists to generate much more data about soil safety. The creators have even made the "recipe" (the code) public, so anyone can use it to help protect our ecosystems.

Note: This tool is specifically designed for counting these specific soil bugs in chemical safety tests. The paper does not claim it can be used for medical diagnoses, counting other types of animals, or predicting future climate scenarios.

Technical Summary: COLLEMBOT – AI-Based Counting of Collembola for OECD 232 Tests

Problem Statement
The assessment of chemical risks in terrestrial ecosystems relies heavily on ecotoxicological tests using soil organisms, specifically the springtail Folsomia candida under the Organization for Economic Co-operation and Development (OECD) 232 reproduction test guidelines. Currently, these protocols depend on the manual counting of juvenile and adult Collembola. This traditional approach presents significant bottlenecks: it is costly, labor-intensive, and time-consuming. Furthermore, manual counting is prone to operator bias, which restricts the availability of data and hinders the robustness of risk assessments.

Methodology
To address these limitations, the authors developed COLLEMBOT, an automated counting tool designed to integrate seamlessly into existing OECD workflows without requiring modifications to the standard protocol. The core of the system is a YOLOv11 convolutional neural network.

Training Data: The model was trained on a dataset of 3,207 high-resolution images.
Validation: The system was rigorously validated using 22 independent datasets comprising 1,704 images. These datasets were sourced from five distinct laboratories across Europe: Amsterdam (Netherlands), Basel (Switzerland), Bayreuth (Germany), Coimbra (Portugal), and Aarhus (Denmark).
Experimental Conditions: The validation covered relevant standard soils, including OECD artificial soils with varying sphagnum peat concentrations (2.5%, 5%, and 10%) and LUFA 2.2 soil, focusing on the detection of Folsomia candida.

Key Results
The performance of COLLEMBOT was evaluated against manual counts across several metrics:

Counting Accuracy: Automated counts demonstrated strong agreement with manual counts, achieving coefficient of determination ( $R^2$ ) values ranging from 0.88 to 0.99.
Toxicological Relevance: Dose-response curves derived from automated counts strongly overlapped with those from manual counts. Effect concentrations (EC values) showed minimal deviation, with a median percentage difference ( $\% \Delta$ ) of 6.2 ± 23. The correlation for EC10 to EC90 values yielded an $R^2 \geq 0.977$ . These results indicate that the automated method remains within acceptable limits for regulatory risk assessment.
Efficiency Gains: The most significant operational improvement was in processing time. A test involving approximately 300 images and up to 1,500 individuals per image was processed in less than 3 hours. In contrast, manual counting for the same volume required approximately 137 hours, representing a 97% reduction in labor time.

Significance and Claims
The paper positions COLLEMBOT as a critical tool for modernizing ecotoxicological testing. By significantly reducing labor requirements and improving reproducibility, the system enables the generation of broader hazard data for Collembola. The authors claim that this advancement directly supports more robust, science-based chemical risk assessments.

To facilitate adoption and foster community-driven development, the authors have made the code and workflow publicly available. The study concludes that COLLEMBOT successfully overcomes the limitations of manual counting while maintaining the integrity required for regulatory compliance.

Technical Summary: COLLEMBOT – AI-Based Counting of Collembola for OECD 232 Tests

More like this