[1Q4-GS-10] AI application:

[1Q4-GS-10-01] Novel features to improve accuracy in body weight estimation of farmed fish

Weight estimation of farmed fish

〇Junya Kobayashi¹, Masashi Tsubaki², Hideki Asoh², Yui Mineshita¹, Ichiro Nagano¹ (1. Nissui Corporation, 2. National Institute of Advanced Industrial Science and Technology)

Keywords:Aquaculture, Body Weight Estimation, Non-contact Measurement, Machine Learning, Feature Exploration

Accurate estimation of the body weight of farmed fish in net pens is crucial for improving the efficiency of aquaculture operations and optimizing their growth. For example, feeding is often determined based on biomass, and precise weight estimation enables the appropriate adjustment of the amount of feed. Non-contact methods minimizing fish stress and damage are preferred for weight estimation. The current mainstream approach for such method involves using fork length and body height measured with underwater stereo cameras as features for weight estimation. However, for achieving higher accuracy, exploration of novel features is required. A dataset of approximately 100,000 samples was constructed, focusing on Japanese amberjack, which includes fork length, body height, body width, the lengths of 11 candidate novel features, and body weight of individuals. Using this dataset, features that could enhance the accuracy of weight estimation models were investigated, and features that significantly contribute to improved accuracy were identified. The body height at 70% of the distance from the snout in the fork length was found to be a critical feature. Incorporating this feature into existing weight estimation models was expected to improve accuracy. In addition, using all significant features was projected to achieve further improvement.