Scholarlink

<p>To address the issue of class imbalance in traditional fishing ground forecasting, this study systematically incorporates seven representative resampling strategies (SMOTE, Borderline-SMOTE, SMOTETomek, SMOTEENN, ADASYN, NearMiss, and random undersampling) combined with ten commonly used machine learning models (<i>k</i>-nearest neighbor [KNN], random forest [RF], gradient boosting decision tree [GBDT], LightGBM, XGBoost, CatBoost, AdaBoost, classification and regression tree [CART], logistic regression [LR], and stacking) to construct an imbalanced classification framework for predicting the albacore tuna (<i>Thunnus alalunga</i>) fishing grounds in the waters near the Cook Islands. Using longline fisheries data and multisource oceanographic variables from 2022 to 2024, the model performance was evaluated through stratified cross-validation and an independent test dataset to examine the classification capability of different resampling model combinations. The results indicate that (1) the imbalanced resampling substantially improves recall and <i>F</i>1-score, effectively mitigating the common overbias toward the majority class (non-primary fishing grounds); (2) among all combinations, the SMOTEENN–LightGBM model achieves the best overall performance (<i>F</i>1 = 0.6805, recall = 0.7886, G-mean = 0.7623), balancing predictive accuracy and computational efficiency; (3) the stacking model demonstrates strong robustness under the non-resampling condition, but its performance gain after resampling is limited; and (4) kernel density similarity analysis shows that the structural similarity index measure (SSIM) value between the predicted and observed primary fishing ground distributions reaches 0.8456, indicating high spatial agreement. These findings suggest that the SMOTEENN–LightGBM model provides a reliable and efficient approach for albacore tuna fishing ground prediction, supporting precise forecasting and sustainable utilization of albacore tuna resources in the waters near the Cook Islands.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Comparative evaluation of imbalanced data processing models for albacore tuna fishing ground prediction: a case study in the waters near Cook Islands

  • 期刊论文

  • 发表日期：

    2026年7月8日

查看全文
  • Liming Song,
  • Chen You,
  • Linhui Wang,
  • Meng Zhang,
  • Shun Wang

摘要

To address the issue of class imbalance in traditional fishing ground forecasting, this study systematically incorporates seven representative resampling strategies (SMOTE, Borderline-SMOTE, SMOTETomek, SMOTEENN, ADASYN, NearMiss, and random undersampling) combined with ten commonly used machine learning models (k-nearest neighbor [KNN], random forest [RF], gradient boosting decision tree [GBDT], LightGBM, XGBoost, CatBoost, AdaBoost, classification and regression tree [CART], logistic regression [LR], and stacking) to construct an imbalanced classification framework for predicting the albacore tuna (Thunnus alalunga) fishing grounds in the waters near the Cook Islands. Using longline fisheries data and multisource oceanographic variables from 2022 to 2024, the model performance was evaluated through stratified cross-validation and an independent test dataset to examine the classification capability of different resampling model combinations. The results indicate that (1) the imbalanced resampling substantially improves recall and F1-score, effectively mitigating the common overbias toward the majority class (non-primary fishing grounds); (2) among all combinations, the SMOTEENN–LightGBM model achieves the best overall performance (F1 = 0.6805, recall = 0.7886, G-mean = 0.7623), balancing predictive accuracy and computational efficiency; (3) the stacking model demonstrates strong robustness under the non-resampling condition, but its performance gain after resampling is limited; and (4) kernel density similarity analysis shows that the structural similarity index measure (SSIM) value between the predicted and observed primary fishing ground distributions reaches 0.8456, indicating high spatial agreement. These findings suggest that the SMOTEENN–LightGBM model provides a reliable and efficient approach for albacore tuna fishing ground prediction, supporting precise forecasting and sustainable utilization of albacore tuna resources in the waters near the Cook Islands.