Monitoring fish growth in commercial aquaculture is traditionally a labor-intensive, often invasive process. Handling fish to manually measure length and weight causes stress, increases the risk of injury or disease, and provides only a snapshot of a population’s health. A team of researchers from Sultan Qaboos University, in collaboration with Abdullah Al Salem University and Oman’s Ministry of Agriculture, Fisheries and Water Resources, has introduced a potential paradigm shift in aquatic resource management: StereoGeoBox.
As detailed in their study, published October 23, 2026, in Computers and Electronics in Agriculture, the researchers have developed a system that estimates fish dimensions and weight without the heavy computational burden of traditional stereo-vision methods. Unlike conventional systems that require generating a complete, dense depth map of an entire underwater scene—a process that is both power-intensive and time-consuming—StereoGeoBox selectively focuses on the fish themselves.
The system utilizes two synchronized cameras paired with YOLO11, a state-of-the-art artificial intelligence model, to identify fish within the camera feeds. Once identified, the system calculates the depth of each fish based on the geometric relationship between the bounding boxes in the paired images. These estimates of length and width are then processed through an XGBoost machine-learning model, which translates the dimensions into an accurate weight prediction.
To validate the technology, the team monitored tilapia over a 15-month period, capturing data across four distinct growth stages. When benchmarked against manual measurements, StereoGeoBox consistently outperformed traditional stereo-vision and learning-based depth methods, producing fewer outliers and achieving higher precision. In a controlled trial involving fish with a reference length of 15.30 centimeters, the system produced an average estimate of 15.25 centimeters, demonstrating a high degree of accuracy.
Beyond accuracy, the system is designed for operational efficiency. By leveraging the NVIDIA Jetson Thor platform, the prototype processes 25.40 frames per second with a latency of just 28.2 milliseconds, using only 17.7 megabytes of GPU memory. This “edge computing” capability, combined with a cloud-based dashboard, allows farm managers to track growth remotely and in real time.
“By estimating depth only where fish are detected, StereoGeoBox reduces computational demands while retaining the speed and accuracy needed for real-time aquaculture monitoring,” explains Dr. Said Al-Abri, the study’s corresponding author.
The researchers also tested the system’s robustness against the harsh realities of underwater environments, including fluctuating illumination, oxygen bubbles, and turbidity. While the system maintained reasonable accuracy under these conditions, the researchers noted that severe visual degradation could limit the number of usable measurements.
Despite these promising results, the team emphasizes that the current iteration of the technology has limitations. The system does not yet facilitate the automatic counting of every individual fish in a tank, nor does it independently calculate an entire farm’s total biomass. Because the system processes individual frames, there is a risk of “double-counting” the same fish as it swims through the field of view.
Future development will focus on scaling the technology and expanding its application to other fish species and diverse commercial aquaculture settings. As the industry looks for more sustainable and efficient ways to feed a growing global population, non-invasive, AI-driven monitoring systems like StereoGeoBox represent a significant step forward in optimizing yields while minimizing the environmental and physical stressors placed on aquatic life.
Disclaimer: This content is auto-generated for informational purposes only.
Source: Read Original News
