Scholarlink

<p>Climate change has significantly altered precipitation patterns in arid and semi- arid regions, presenting substantial challenges for water resource management and agricultural planning. This study assesses the effectiveness of machine learning (ML) models, specifically Long Short-Term Memory (LSTM) networks, for precipitation forecasting in Lubbock, Texas, a semi- arid region characterized by high variability and climatic uncertainty. Utilizing meteorological data from 2012 to 2023, which includes 82,475 daily observations, we systematically compare traditional statistical methods (linear regression, ARIMA, SARIMA) with advanced ML techniques such as Multi-Layer Perceptron (MLP) and LSTM. The traditional approaches exhibit limited performance, with linear regression achieving an R ² of 0.512 and SARIMA an MSE of 0.178, owing to their inability to capture non-linear and temporal dependencies. To overcome this limitation, we developed and refined LSTM networks, integrating region-specific adaptations such as zero-inflation handling (73% non-precipitation days) and the incorporation of mixed precipitation phenomena (e.g., graupel, hail). Our methodology demonstrates notable enhancements, with the LSTM model achieving a Mean Squared Error (MSE) of 0.0218 and an R ² of 0. 957, thereby displaying its capacity to forecast both typical precipitation events and rare phenomena, such as a graupel event (0. 001 inches) with 89% accuracy. Methodological innovations include the implementation of an extended LSTM architecture for forecasting with uncertainty bounds of ± 0. 298 inches, optimized for zero- inflated data, as well as the integration of diverse precipitation phenomena and a stakeholder- focused validation framework. The findings offer a comprehensive, transferable framework for improving water resource management, agricultural planning, and emergency preparedness in arid regions worldwide, with estimated economic benefits exceeding $ 34.5&#xa0;million annually for the study area. This research establishes a replicable framework for ML-enhanced precipitation forecasting in arid regions, addressing critical gaps in climate adaptation strategies for semi-arid territories.</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

Machine learning-enhanced precipitation forecasting in arid regions: advancing climate modelling techniques in Lubbock, Texas

  • 期刊论文

  • 发表日期：

    2026年4月18日

查看全文
  • M Shahriar Sonet,
  • Md Yeasir Hasan,
  • Abdulla Al Kafy,
  • Hamad Ahmed Altuwaijri

摘要

Climate change has significantly altered precipitation patterns in arid and semi- arid regions, presenting substantial challenges for water resource management and agricultural planning. This study assesses the effectiveness of machine learning (ML) models, specifically Long Short-Term Memory (LSTM) networks, for precipitation forecasting in Lubbock, Texas, a semi- arid region characterized by high variability and climatic uncertainty. Utilizing meteorological data from 2012 to 2023, which includes 82,475 daily observations, we systematically compare traditional statistical methods (linear regression, ARIMA, SARIMA) with advanced ML techniques such as Multi-Layer Perceptron (MLP) and LSTM. The traditional approaches exhibit limited performance, with linear regression achieving an R ² of 0.512 and SARIMA an MSE of 0.178, owing to their inability to capture non-linear and temporal dependencies. To overcome this limitation, we developed and refined LSTM networks, integrating region-specific adaptations such as zero-inflation handling (73% non-precipitation days) and the incorporation of mixed precipitation phenomena (e.g., graupel, hail). Our methodology demonstrates notable enhancements, with the LSTM model achieving a Mean Squared Error (MSE) of 0.0218 and an R ² of 0. 957, thereby displaying its capacity to forecast both typical precipitation events and rare phenomena, such as a graupel event (0. 001 inches) with 89% accuracy. Methodological innovations include the implementation of an extended LSTM architecture for forecasting with uncertainty bounds of ± 0. 298 inches, optimized for zero- inflated data, as well as the integration of diverse precipitation phenomena and a stakeholder- focused validation framework. The findings offer a comprehensive, transferable framework for improving water resource management, agricultural planning, and emergency preparedness in arid regions worldwide, with estimated economic benefits exceeding $ 34.5 million annually for the study area. This research establishes a replicable framework for ML-enhanced precipitation forecasting in arid regions, addressing critical gaps in climate adaptation strategies for semi-arid territories.