Introduction
Accurate biomass estimation is crucial for effective fish farm management, facilitating optimized feeding, health assessment, disease control, and resource allocation. Manual counting and weighing are labor-intensive, leading to the rise of computer vision solutions powered by artificial intelligence. However, hardware costs and computing power remain limiting factors, particularly for three-dimensional information requirements. While advanced methods like 3D-image synthesis and stereo-vision cameras are effective for intensive industrial farms, a more accessible solution is needed for widespread adoption in the aquaculture industry. In collaboration, the Zurich University of Applied Sciences and Urban Blue have initiated the AWACS (Animal Welfare Assessment and Control System) innovation project, funded by the Innosuisse agency. The aim is to develop an automated assessment system for fish farms, enabling continuous monitoring, automatic assessment, and visual analysis of fish health and welfare.
Biomass modeling in a 3D environment
A biomass model using computer vision was developed to calculate fish weight based on single frames extracted from underwater camera videos. The third-dimensional information was derived from the camera’s focal length. Increasing the camera’s aperture reduces the depth of field, causing objects outside a certain distance to appear blurry. Object detection identifies fish in the frame and determines their blur value, excluding fish outside the focus area and retaining only those at a known horizontal distance from the camera lens . The reduction of data from a video to a subsample of several hundred high-quality single frames with fish suitable for the biomass model increases the quality of the biomass estimation and reduces the computing power and digital storage necessary.
Test results show the applicability of the algorithm
During purging, a 10 m3 RAS tank with 850 rainbow trouts was videotaped the day before slaughtering , and the weight was taken of each fish at slaughtering. The biomass model was calibrated with lateral pictures of 40 of these fish, and the video was analyzed with the algorithm. The estimated body weights were compared with the weights measured with the scale, and preliminary analysis shows a good fit of the two datasets, indicating the functionality of the method. This low-cost, low-computation, and easy-to-install setup improves real-time biomass estimation in aquaculture, both in research and industrial settings.