Scholarlink

<p>Excessive algal growth poses serious threats to aquatic ecosystems and water safety. Chlorophyll-<i>a</i> (Chl-<i>a</i>) serves as a key indicator of algal presence in rivers and lakes. Although various machine learning (ML) models have been widely applied to algal bloom prediction, field measurement data are often imbalanced, which can adversely affect the performance of ML models. This study employed a one-dimensional (1D) convolutional neural network (CNN) combined with a long short-term memory (LSTM) network to classify algal bloom levels into three classes based on Chl-<i>a</i> concentrations. To address class imbalance in the dataset, the synthetic minority over-sampling technique combined with the edited nearest neighbors (SMOTEENN) was applied. SMOTEENN combines SMOTE for over-sampling the minority classes with ENN for under-sampling and data cleaning. By addressing data imbalance and removing noise near the decision boundary, the application of SMOTEENN improved the data distribution balance. The model performance improved significantly, with the macro-averages of precision, recall, and F1 score increasing from 0.44, 0.50, and 0.47 using the original data to 0.74, 0.69, and 0.70, respectively, after applying SMOTEENN. Notable improvements were achieved for Class 3, with its sample size increasing from 1,505 to 12,731 by applying SMOTEENN, making it comparable to the other classes. The model using the original data failed to classify Class 3, whereas the SMOTEENN-applied model achieved significantly improved precision, recall, and F1 scores of 0.80, 0.62, and 0.68, respectively. This study demonstrates that addressing data imbalance can significantly enhance deep learning model performance, improving its applicability to field management.</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

Improvement of deep learning model performance for algal bloom prediction by resolving data imbalance in field observations

  • 期刊论文

  • 发表日期：

    2025年7月11日

查看全文
  • Juneoh Kim,
  • Woo Hyoung Lee,
  • Jungsu Park

摘要

Excessive algal growth poses serious threats to aquatic ecosystems and water safety. Chlorophyll-a (Chl-a) serves as a key indicator of algal presence in rivers and lakes. Although various machine learning (ML) models have been widely applied to algal bloom prediction, field measurement data are often imbalanced, which can adversely affect the performance of ML models. This study employed a one-dimensional (1D) convolutional neural network (CNN) combined with a long short-term memory (LSTM) network to classify algal bloom levels into three classes based on Chl-a concentrations. To address class imbalance in the dataset, the synthetic minority over-sampling technique combined with the edited nearest neighbors (SMOTEENN) was applied. SMOTEENN combines SMOTE for over-sampling the minority classes with ENN for under-sampling and data cleaning. By addressing data imbalance and removing noise near the decision boundary, the application of SMOTEENN improved the data distribution balance. The model performance improved significantly, with the macro-averages of precision, recall, and F1 score increasing from 0.44, 0.50, and 0.47 using the original data to 0.74, 0.69, and 0.70, respectively, after applying SMOTEENN. Notable improvements were achieved for Class 3, with its sample size increasing from 1,505 to 12,731 by applying SMOTEENN, making it comparable to the other classes. The model using the original data failed to classify Class 3, whereas the SMOTEENN-applied model achieved significantly improved precision, recall, and F1 scores of 0.80, 0.62, and 0.68, respectively. This study demonstrates that addressing data imbalance can significantly enhance deep learning model performance, improving its applicability to field management.