Chaitanya Baliram Pande, Dinesh Kumar Vishwakarma, Aman Srivastava, Kanak N. Moharir, Fahad Alshehri, Norashidah Md Din, Lariyah Mohd Sidek, Bojan Đurin, Abebe Debele Tolche
{"title":"半干旱气候区气象干旱预报的机器学习模型","authors":"Chaitanya Baliram Pande, Dinesh Kumar Vishwakarma, Aman Srivastava, Kanak N. Moharir, Fahad Alshehri, Norashidah Md Din, Lariyah Mohd Sidek, Bojan Đurin, Abebe Debele Tolche","doi":"10.1007/s13201-025-02445-x","DOIUrl":null,"url":null,"abstract":"<div><p>The central region of Maharashtra, India, is susceptible to agriculture, meteorological, and hydrological droughts, impacting local ecosystems. The scarcity of historical data impedes monitoring and forecasting regional droughts. Given the limited studies on ensemble and Machine Learning (ML) models for drought forecasting, this research compares five ML models [Robust Linear Regression, Bagged Trees, Boosted Trees, Support Vector Machine (SVM), and Matern Gaussian Process Regression (GPR)] to determine superior accuracy in the regional context. The study aims to assess the accuracy of the developed models in predicting future drought events and gain insights into meteorological droughts in tropical climates using the Standardized Precipitation Index (SPI)-SPI-3 and SPI-6. Subset regression analysis exhibited SPI-1, -3, -4, -5, and -6 as the best input subsets for SPI-3, whereas SPI-1 and -2 for SPI-6. Results indicated that the Matern GPR model outperformed other models in SPI-3 and SPI-6 training phases (MSE = 0.1954, 0.0493; RMSE = 0.4420, 0.2221; MAE = 0.3382, 0.1683; MARE = 1.3807, 0.5237; NSE = 0.6585, 0.9048; R = 0.9165, 0.9920; R<sup>2</sup> = 0.8399, 0.9841). In testing, the SVM model bettered in SPI-3 and SPI-6 forecasting (MSE = 0.5735, 0.8479; RMSE = 0.7573, 0.9208; MAE = 0.5882, 0.5300; MARE = − 0.5638, − 0.3621; NSE = 0.8676, 0.8601; R = 0.9317, 0.9275; R<sup>2</sup> = 0.8680, 0.8603). The ensemble method played a novel and crucial role in significantly improving the accuracy of drought forecasting by developing ML models based on various algorithms that operate more efficiently, require fewer inputs, and exhibit less complexity than precise models, proving highly effective for drought warning systems. Therefore, results offer valuable insights for crop planning, drought challenges, water management, and maintaining the study area ecosystem. In conclusion, the study addressed the challenge proposed by incomplete previous data for monitoring and forecasting regional drought events by employing advanced data imputation techniques, ensemble learning methods, and incorporating robust ML models like SVM and Matern GPR.</p></div>","PeriodicalId":8374,"journal":{"name":"Applied Water Science","volume":"15 6","pages":""},"PeriodicalIF":5.7000,"publicationDate":"2025-05-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://link.springer.com/content/pdf/10.1007/s13201-025-02445-x.pdf","citationCount":"0","resultStr":"{\"title\":\"A novel machine learning models for meteorological drought forecasting in the semi-arid climate region\",\"authors\":\"Chaitanya Baliram Pande, Dinesh Kumar Vishwakarma, Aman Srivastava, Kanak N. Moharir, Fahad Alshehri, Norashidah Md Din, Lariyah Mohd Sidek, Bojan Đurin, Abebe Debele Tolche\",\"doi\":\"10.1007/s13201-025-02445-x\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><p>The central region of Maharashtra, India, is susceptible to agriculture, meteorological, and hydrological droughts, impacting local ecosystems. The scarcity of historical data impedes monitoring and forecasting regional droughts. Given the limited studies on ensemble and Machine Learning (ML) models for drought forecasting, this research compares five ML models [Robust Linear Regression, Bagged Trees, Boosted Trees, Support Vector Machine (SVM), and Matern Gaussian Process Regression (GPR)] to determine superior accuracy in the regional context. The study aims to assess the accuracy of the developed models in predicting future drought events and gain insights into meteorological droughts in tropical climates using the Standardized Precipitation Index (SPI)-SPI-3 and SPI-6. Subset regression analysis exhibited SPI-1, -3, -4, -5, and -6 as the best input subsets for SPI-3, whereas SPI-1 and -2 for SPI-6. Results indicated that the Matern GPR model outperformed other models in SPI-3 and SPI-6 training phases (MSE = 0.1954, 0.0493; RMSE = 0.4420, 0.2221; MAE = 0.3382, 0.1683; MARE = 1.3807, 0.5237; NSE = 0.6585, 0.9048; R = 0.9165, 0.9920; R<sup>2</sup> = 0.8399, 0.9841). In testing, the SVM model bettered in SPI-3 and SPI-6 forecasting (MSE = 0.5735, 0.8479; RMSE = 0.7573, 0.9208; MAE = 0.5882, 0.5300; MARE = − 0.5638, − 0.3621; NSE = 0.8676, 0.8601; R = 0.9317, 0.9275; R<sup>2</sup> = 0.8680, 0.8603). The ensemble method played a novel and crucial role in significantly improving the accuracy of drought forecasting by developing ML models based on various algorithms that operate more efficiently, require fewer inputs, and exhibit less complexity than precise models, proving highly effective for drought warning systems. Therefore, results offer valuable insights for crop planning, drought challenges, water management, and maintaining the study area ecosystem. In conclusion, the study addressed the challenge proposed by incomplete previous data for monitoring and forecasting regional drought events by employing advanced data imputation techniques, ensemble learning methods, and incorporating robust ML models like SVM and Matern GPR.</p></div>\",\"PeriodicalId\":8374,\"journal\":{\"name\":\"Applied Water Science\",\"volume\":\"15 6\",\"pages\":\"\"},\"PeriodicalIF\":5.7000,\"publicationDate\":\"2025-05-05\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"https://link.springer.com/content/pdf/10.1007/s13201-025-02445-x.pdf\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Applied Water Science\",\"FirstCategoryId\":\"93\",\"ListUrlMain\":\"https://link.springer.com/article/10.1007/s13201-025-02445-x\",\"RegionNum\":3,\"RegionCategory\":\"环境科学与生态学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q1\",\"JCRName\":\"WATER RESOURCES\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Applied Water Science","FirstCategoryId":"93","ListUrlMain":"https://link.springer.com/article/10.1007/s13201-025-02445-x","RegionNum":3,"RegionCategory":"环境科学与生态学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"WATER RESOURCES","Score":null,"Total":0}
A novel machine learning models for meteorological drought forecasting in the semi-arid climate region
The central region of Maharashtra, India, is susceptible to agriculture, meteorological, and hydrological droughts, impacting local ecosystems. The scarcity of historical data impedes monitoring and forecasting regional droughts. Given the limited studies on ensemble and Machine Learning (ML) models for drought forecasting, this research compares five ML models [Robust Linear Regression, Bagged Trees, Boosted Trees, Support Vector Machine (SVM), and Matern Gaussian Process Regression (GPR)] to determine superior accuracy in the regional context. The study aims to assess the accuracy of the developed models in predicting future drought events and gain insights into meteorological droughts in tropical climates using the Standardized Precipitation Index (SPI)-SPI-3 and SPI-6. Subset regression analysis exhibited SPI-1, -3, -4, -5, and -6 as the best input subsets for SPI-3, whereas SPI-1 and -2 for SPI-6. Results indicated that the Matern GPR model outperformed other models in SPI-3 and SPI-6 training phases (MSE = 0.1954, 0.0493; RMSE = 0.4420, 0.2221; MAE = 0.3382, 0.1683; MARE = 1.3807, 0.5237; NSE = 0.6585, 0.9048; R = 0.9165, 0.9920; R2 = 0.8399, 0.9841). In testing, the SVM model bettered in SPI-3 and SPI-6 forecasting (MSE = 0.5735, 0.8479; RMSE = 0.7573, 0.9208; MAE = 0.5882, 0.5300; MARE = − 0.5638, − 0.3621; NSE = 0.8676, 0.8601; R = 0.9317, 0.9275; R2 = 0.8680, 0.8603). The ensemble method played a novel and crucial role in significantly improving the accuracy of drought forecasting by developing ML models based on various algorithms that operate more efficiently, require fewer inputs, and exhibit less complexity than precise models, proving highly effective for drought warning systems. Therefore, results offer valuable insights for crop planning, drought challenges, water management, and maintaining the study area ecosystem. In conclusion, the study addressed the challenge proposed by incomplete previous data for monitoring and forecasting regional drought events by employing advanced data imputation techniques, ensemble learning methods, and incorporating robust ML models like SVM and Matern GPR.