Artificial intelligence analysis for feed ratio optimization in precision aquaculture

• Main Author: Galenius, Bryan Timothy
• Other Authors: Ng Yin Kwee
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Engineering; Aquaculture; Computer vision; Satiety; Machine learning; Hunger detection
• Online Access: https://hdl.handle.net/10356/176594
• Institution: Nanyang Technological University

Aquaculture products’ increasing demand emphasizes the importance of FCR (Feed Conversion Ratio) optimization. While efforts to improve FCR of specific species exist, a universal method is lacking. Therefore, we present a hunger detection method generalized for different aquaculture species by summarizing and combining the different behavioral parameters which has been found to be useful in determining hunger into 5 parameters: Weighted Positional Index, Position Variance, Average Speed, Lingering Count, and Average Swim Direction and Turning Angle. We continuously recorded 21 Danio rerio (Zebrafish), a fish which has yet to be tested in hunger detection, in a video for 7 days twice. The first 7-day recording is done with a twice-a-day feeding regime, while the second recording is with only once-a-day feeding regime. The recordings are then passed through DETR object detection algorithm and StrongSORT multi-object tracking algorithm to extract the 5 proposed parameters, after which each parameter’s role in hunger detection is evaluated through the comparison between the 2 different feeding regimes. Finally, the twice-a-day feeding regime data is passed onto multiple different Machine Learning (ML) algorithm. If the algorithm manages to find one or more clusters where all the values inside have a high value of ‘time elapsed since the last feeding’ and with all clusters having a relatively large number of datapoints each cluster (indicating the absence overfitting), then we can conclude that the 5 parameters are indeed enough for the algorithm to find a cluster in which the fish is hungry. The result shows that the proposed algorithm pipeline was indeed able to utilize the 5 proposed parameters to identify a cluster showing fish hunger through Gaussian Mixture Model clustering on PCA-transformed parameter data. It also shows that each parameter indeed contributes to the hunger prediction. Therefore, the proposed algorithm pipeline can move forward to be tested on multiple different species to ensure the generalization of the algorithm. Any other relevant parameters found in the future can also be easily added. Afterwards, it can be further integrated into the hardware system to create a holistic smart feeding system usable in aquaculture farms.