Automated prediction of spawning nights using machine learning analysis of flatfish behaviour

Identifier:  imarina:9470396

Handle: https://hdl.handle.net/20.500.11797/imarina9470396

Authors:  Qadir, A; Duncan, N; González-Lopez, WA; Fatsini, E; Serratosa, F

Abstract:
Senegalese sole (Solea senegalensis) broodstock exhibit distinct behaviours (Rest the Head, Guardian, Follow, and Locomotor activities) that are important for breeding success. Understanding and monitoring these behaviours are essential to understand successful breeding of Senegalese sole. However, manually analysing these behaviours represents a significant challenge for human observers and is a labour-intensive process. Moreover, due to reproductive dysfunctions in Senegalese sole, aquaculture operations currently depend on wild origin breeders for successful spawning a reliance that is unsustainable in the long term. Therefore, to address these limitations, this study introduces a custom-designed framework based on computer vision and machine learning techniques. The model integrates object detection and tracking mechanisms to recognize and monitor reproductive behaviours of Senegalese sole within aquaculture environments. By combining advanced tracking algorithms, our model effectively extracts and analyses behavioural patterns from video datasets. The automated model behavioural analyses compared with manual analyses demonstrated strong performance, with accuracy, precision, and specificity exceeding 87 %, and a Pearson correlation of R = 0.99 between manual observation data and automated data. The model analysed videos to accurately identify behaviours with minimal human intervention, thereby saving a substantial number of hours and opened up the possibility to analyse behaviours over longer periods, generating more data. This is the first study to automatically analyse reproductive behaviours across full-night video recordings in Senegalese sole, providing new insights into how behavioural patterns relate to spawning. These behavioural changes in relation to spawning enable the model to effectively predict spawning and non-spawning nights with accuracies ranging from 70 % to 100 %. Such predictive capability can reduce dependence on wild origin breeders, support timely gamete collection, improve reproductive planning, and serve as a potential tool for hatchery automation.