Decision Tree-based Prediction of Hatching Efficiency in Mud Spiny Lobster (Panulirus polyphagus): Relative Importance of Body Weight, Carapace Length, and Water Quality

Abstract

Machine learning (ML) has been applied increasingly in aquaculture; however, its practical use in hatchery management remains limited due to challenges in model interpretability. This study developed an interpretable decision tree (DT) model to predict the hatching efficiency (HE) of mud spiny lobster (Panulirus polyphagus) in a hatchery system (n = 72), considering broodstock size indicators—carapace length (CL) and body weight (BW)—and key water quality parameters—temperature, salinity, dissolved oxygen (DO), pH, total ammonia nitrogen (TAN), nitrite (NO₂^-), and alkalinity). The objectives were to compare BW and CL performance, identify key factors affecting hatching success, and establish threshold-based decision rules for hatchery management. Preliminary regression analysis showed that simple linear models could not adequately capture the relationship between broodstock size and HE. However, BW performed relatively better than CL and was therefore selected for subsequent DT modeling. HE was classified into two groups using K-means clustering: low (<40.5%) and high (≥40.5%). The DT model achieved an overall classification accuracy of 62.50% based on 10-fold cross-validation of the complete dataset (n = 72), with variable classification performance between HE groups. In particular, the model showed limited ability to identify the high-HE group, with a recall of 0.11 and F1-score of 0.18. TAN, DO, and BW were identified as the most influential variables, with nitrite also contributing to the decision rules for hatchery management. Maintaining TAN below 0.14 mg-N·L^-1 was associated with higher HE, particularly in broodstock weighing 819.5–895.6 g, while a higher TAN value was associated with reduced performance. In addition, DO levels above 4.95 mg·L^-1 improved HE in larger broodstock (>904 g). Based on these findings, the DT model may serve as an exploratory and interpretable decision-support tool for understanding HE patterns and informing broodstock and water quality management in sustainable lobster hatchery systems, while further model development and validation are needed before its use as a reliable predictive system.