Scholarlink

<p>Drought is a major natural hazard that disrupts agriculture, water resources, and ecosystems. Reliable, multi-timescale drought forecasting is essential for effective water management; however, the non-stationary and non-linear nature of drought data makes accurate predictions challenging. This study presents a regression-based approach that combines variational mode decomposition (VMD) and metaheuristic optimization to enhance multi-step drought predictions. Using Standardized Precipitation Evapotranspiration Index (SPEI<sub>12</sub>) data, we compare three machine learning algorithms—Extra Trees Regression (ET), K-Nearest Neighbors (KNN), and Kernel Ridge Regression (KRR)—with and without data decomposition. Our analysis focuses on three Atlantic Canadian locations (Charlottetown, Saint John, and Sydney), where rainfall-dependent agriculture may be vulnerable to drought. The application of autocorrelation and partial autocorrelation functions determined effective input lags; the KRR algorithm was optimized using Sand Cat Swarm Optimization (SCSO). Results show that the VMD-KRR-SCSO model delivers the most accurate 3- and 6-month drought predictions at all research sites. For one month ahead, it achieved exceptional accuracy, with correlation coefficient (R) values of 0.9949 in Charlottetown, 0.9930 in Saint John, and 0.9884 in Sydney, and low root mean square error (RMSE) values of 0.1100, 0.1093, and 0.1422, respectively. While accuracy declined at longer forecast horizons, the model consistently outperformed single-model approaches. These findings highlight VMD-KRR-SCSO as a powerful tool for improving drought prediction, offering valuable insights for early warning systems and sustainable drought 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

Improving multi-step drought forecasting in Atlantic Canada through variational mode decomposition and machine learning: The role of sand-cat swarm optimization technique in kernel ridge regression

  • 期刊论文

  • 发表日期：

    2026年3月13日

查看全文
  • Chukwuemeka Eneh,
  • Gurjit S. Randhawa,
  • Masoud Karbasi,
  • Aitazaz Ahsan Farooque,
  • Mumtaz Ali

摘要

Drought is a major natural hazard that disrupts agriculture, water resources, and ecosystems. Reliable, multi-timescale drought forecasting is essential for effective water management; however, the non-stationary and non-linear nature of drought data makes accurate predictions challenging. This study presents a regression-based approach that combines variational mode decomposition (VMD) and metaheuristic optimization to enhance multi-step drought predictions. Using Standardized Precipitation Evapotranspiration Index (SPEI12) data, we compare three machine learning algorithms—Extra Trees Regression (ET), K-Nearest Neighbors (KNN), and Kernel Ridge Regression (KRR)—with and without data decomposition. Our analysis focuses on three Atlantic Canadian locations (Charlottetown, Saint John, and Sydney), where rainfall-dependent agriculture may be vulnerable to drought. The application of autocorrelation and partial autocorrelation functions determined effective input lags; the KRR algorithm was optimized using Sand Cat Swarm Optimization (SCSO). Results show that the VMD-KRR-SCSO model delivers the most accurate 3- and 6-month drought predictions at all research sites. For one month ahead, it achieved exceptional accuracy, with correlation coefficient (R) values of 0.9949 in Charlottetown, 0.9930 in Saint John, and 0.9884 in Sydney, and low root mean square error (RMSE) values of 0.1100, 0.1093, and 0.1422, respectively. While accuracy declined at longer forecast horizons, the model consistently outperformed single-model approaches. These findings highlight VMD-KRR-SCSO as a powerful tool for improving drought prediction, offering valuable insights for early warning systems and sustainable drought management.