Swimming crab (Portunus trituberculatus) is a commercially valuable species in the Yellow Sea, where recent fluctuations in resource levels have raised concerns about sustainable management. This study aimed to improve the estimation of the carapace length at 50% maturity (L50) using machine learning techniques, providing a more consistent and reproducible framework for visual maturity classification by standardizing image-based decision processes. Using geometric image augmentation (e.g., rotation, flipping, brightness adjustment), Hue–Saturation–Value (HSV) color segmentation, and algorithms, such as Extreme Gradient Boosting (XGB), Support Vector Machine (SVM), Random Forest (RF), and ensemble models, we classified the maturity of female crabs based on gonad color features. Model performance was evaluated using accuracy, AUC, and the TSS, with the ensemble model showing the highest predictive capability. The machine learning-based L50 was estimated at 64.63 mm (±1.73 mm), yielding a narrower uncertainty range than the visually derived L50 of 65.47 mm (±2.89 mm) under the same macroscopic labeling framework. These results suggest that machine learning-assisted maturity classification can enhance the precision and operational consistency of maturity estimation under a standardized framework, while biological accuracy cannot be confirmed in the absence of an independent reference, such as histological validation.
Swimming crab (Portunus trituberculatus) is a commercially valuable species in the Yellow Sea, where recent fluctuations in resource levels have raised concerns about sustainable management. This study aimed to enhance the estimation of the carapace length at 50% maturity (L₅₀) through machine learning techniques, offering a more objective alternative to traditional visual inspection. Using geometric image augmentation (e.g., rotation, flipping, brightness adjustment), Hue-Saturation-Value (HSV) color segmentation, and algorithms such as Extreme Gradient Boosting (XGB), Support Vector Machine (SVM), Random Forest (RF), and ensemble models, we classified the maturity of female crabs based on gonad color features. Model performance was evaluated using Accuracy, AUC, and TSS, with the ensemble model showing the highest predictive capability. The machine learning-based L₅₀ was estimated at 64.63 mm (±1.73 mm), which was more precise than the visually derived L₅₀ of 65.47 mm (±2.89 mm). These results suggest that machine learning techniques can serve as reliable tools for developing science-based management strategies, ultimately supporting sustainable fisheries resource management.