Scholarlink

<p>Effective drought prediction plays a vital role in lessening the effects of water scarcity on various sectors, including water resource management, agricultural practices, and ecological systems. Machine learning (ML) models present a potentially valuable tool for drought forecasting because of their computational advantages and streamlined architectures. This study investigated the efficacy of ML models for drought forecasting in Iran’s Persian Gulf and Caspian Sea coastal regions, characterized by extreme climates. Focusing on the Standardized Precipitation Evapotranspiration Index (SPEI) as a drought indicator, the research evaluated four ML algorithms: ANN, MARS, RF, and BART. Utilizing historical data from 1966 to 2023, the study employed the Uncertainty Estimation based on local Errors and Clustering (UNEEC) method to quantify model uncertainty. Performance metrics, including RMSE, MAE, and R², revealed the superior performance of ANN and MARS. These models achieved the lowest RMSE (0.23 for ANN) and MAE (0.18 for MARS) values, coupled with the highest R² (0.92 for both). This suggests a stronger ability of ANN and MARS to model the complex relationships influencing SPEI compared to RF and BART. Furthermore, the application of UNEEC demonstrated higher Prediction Interval Coverage Probability (PICP), averaging 0.91, for ANN and MARS, indicating greater prediction reliability. This research introduces a novel approach by integrating multiple ML models with the UNEEC method to conduct a thorough uncertainty analysis. This combined approach facilitates robust model comparison and evaluation of prediction reliability. These results provide important information that can enhance long-term water resource strategies, inform agricultural planning, and improve drought mitigation efforts in Iran’s coastal areas. The improved SPEI predictions generated by this research can support more effective decision-making processes.</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

Uncertainty analysis of machine learning models for SPEI estimation in the Persian Gulf and Caspian Sea

  • 期刊论文

  • 发表日期：

    2025年7月23日

查看全文
  • Mojtaba Mohammadi,
  • Ommolbanin Bazrafshan,
  • Hossein Zamani

摘要

Effective drought prediction plays a vital role in lessening the effects of water scarcity on various sectors, including water resource management, agricultural practices, and ecological systems. Machine learning (ML) models present a potentially valuable tool for drought forecasting because of their computational advantages and streamlined architectures. This study investigated the efficacy of ML models for drought forecasting in Iran’s Persian Gulf and Caspian Sea coastal regions, characterized by extreme climates. Focusing on the Standardized Precipitation Evapotranspiration Index (SPEI) as a drought indicator, the research evaluated four ML algorithms: ANN, MARS, RF, and BART. Utilizing historical data from 1966 to 2023, the study employed the Uncertainty Estimation based on local Errors and Clustering (UNEEC) method to quantify model uncertainty. Performance metrics, including RMSE, MAE, and R², revealed the superior performance of ANN and MARS. These models achieved the lowest RMSE (0.23 for ANN) and MAE (0.18 for MARS) values, coupled with the highest R² (0.92 for both). This suggests a stronger ability of ANN and MARS to model the complex relationships influencing SPEI compared to RF and BART. Furthermore, the application of UNEEC demonstrated higher Prediction Interval Coverage Probability (PICP), averaging 0.91, for ANN and MARS, indicating greater prediction reliability. This research introduces a novel approach by integrating multiple ML models with the UNEEC method to conduct a thorough uncertainty analysis. This combined approach facilitates robust model comparison and evaluation of prediction reliability. These results provide important information that can enhance long-term water resource strategies, inform agricultural planning, and improve drought mitigation efforts in Iran’s coastal areas. The improved SPEI predictions generated by this research can support more effective decision-making processes.