AI-Powered Biomass Monitoring – Automating Underwater Life Tracking

The client is a leading German research institution focused on polar, climate, and marine sciences. Its Kiel-based division specializes in innovative solutions for sustainable aquaculture. One of their flagship projects involved industrial-scale shrimp farming, during which the institute conducted experiments aimed at optimizing production processes and monitoring animal health. In collaboration with the research team in Kiel, we initiated the development of a system that enables automatic, non-invasive shrimp counting and assessment of biomass dynamics.

One of the main challenges faced by the institute was the need for regular monitoring of shrimp population and biomass changes. These data are essential for evaluating farming quality, planning sales, and making informed production decisions. Until now, this information was collected manually through sampling — a method that was not only time-consuming and imprecise but also stressful and potentially harmful to the animals.

The goal was to develop a system that could automatically count shrimp without the need for extraction. Key requirements included high measurement accuracy, resilience to variable conditions (such as lighting, shrimp density and color, and camera angles), and the ability to perform continuous monitoring without disrupting farm operations.

Stage 1: Prototype and Rapid Start

To quickly validate the concept, we built an initial system using an iPhone connected to a constant power source, capturing images every minute. These images were automatically uploaded to a server. Despite its temporary nature, the setup enabled early validation of data quality and revealed infrastructure limitations—such as unstable Wi-Fi in the aquaculture facility.

Stage 2: Infrastructure Scale-Up

Following a successful prototype, we delivered a full suite of professional-grade imaging equipment and deployed a computing environment (servers) capable of continuous and stable data processing. The entire system architecture was tailored to the needs of Computer Vision (CV).

Stage 3: AI Model Development and Testing

We began building the AI module responsible for image analysis:

1. A diversified image dataset from the shrimp farm was prepared and thoroughly annotated (including various lighting conditions, densities, camera angles, and shrimp colors).
2. We tested three model architectures: Faster R-CNN (two-stage detector), YOLOv5 (single-stage detector), and a density map-based U2-Net model.
3. Custom neural network layers were developed, and aggressive data augmentation was applied to improve performance under challenging conditions—such as high shrimp density.

Stage 4: Validation and Optimization

The models were tested on a held-out validation set, including out-of-distribution data.

• YOLOv5 delivered the best results, achieving a relative counting error of approximately 6%.
• The system successfully handled scenes with over 200 shrimp per frame.

Stage 5: Visualization and Data Accessibility

To provide real-time insights to the research team, we developed a lightweight web application using Streamlit. It enables real-time result viewing and access to the image analysis archive.

• Continuous farm monitoring – image capture every minute with real-time analysis
• Accurate shrimp count and biomass data – improved production and sales management
• Reduced stress and injury to shrimp – no need for physical extraction
• Automation of manual, labor-intensive tasks – saving time and human resources
• Scalability potential – system can be adapted for other species (salmon, lobsters, mussels)
• Positive reception and commercialization plans – client chose to expand the solution beyond the PoC phase