Scholarlink

<p>Accurate reference evapotranspiration (ET<sub>O</sub>) estimation with limited meteorological inputs and consistent performance across regions is a crucial for water resource management, highlighting the need for more robust and transferable approaches. This study utilizes nine empirical equations across three categories (temperature-based models, mass transfer-based models, and radiation-based models) and six machine learning (ML) models (including tree-based, kernel-based and neural network-based) to estimate ET<sub>O</sub> at the stations Pusa and Nagpur in India. The empirical equations from each category and the six ML models, which utilize all inputs associated with the corresponding empirical equations are compared against the standard FAO-56 Penman–Monteith estimate to assess their accuracy in estimating ET<sub>O</sub>. Among the empirical models, radiation-based Ritchie's estimates are superior, with RMSE values of 0.493&#xa0;mm/day at Pusa. Among the ML models with radiation-based inputs, the SVR model performs the best at Pusa station, with RMSE of 0.345&#xa0;mm/day. Considering the spatial transferability, the temperature-based CNN model developed for Pusa exhibited the highest transferability score of 1.0. The best performance by an ML model beyond the study station is observed for the radiation-based SVR model from Pusa: R<sup>2</sup> = 0.895, d = 0.945, RMSE = 0.87&#xa0;mm/day, MAPE = 0.150. These findings indicate that the application of ML models can significantly enhance the accuracy of ET<sub>O</sub> estimation while achieving effective transferability compared to empirical models. By addressing the limitations of traditional approaches, the methodology developed in this study can be applied in other regions to estimate ET<sub>O</sub> and to model other complex hydrological 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

Empirical and ML models for estimation of evapotranspiration: comparison of performance and evaluation of spatial transferability

  • 期刊论文

  • 发表日期：

    2026年2月7日

查看全文
  • Nehar Mandal,
  • Thendiyath Roshni,
  • Kironmala Chanda

摘要

Accurate reference evapotranspiration (ETO) estimation with limited meteorological inputs and consistent performance across regions is a crucial for water resource management, highlighting the need for more robust and transferable approaches. This study utilizes nine empirical equations across three categories (temperature-based models, mass transfer-based models, and radiation-based models) and six machine learning (ML) models (including tree-based, kernel-based and neural network-based) to estimate ETO at the stations Pusa and Nagpur in India. The empirical equations from each category and the six ML models, which utilize all inputs associated with the corresponding empirical equations are compared against the standard FAO-56 Penman–Monteith estimate to assess their accuracy in estimating ETO. Among the empirical models, radiation-based Ritchie's estimates are superior, with RMSE values of 0.493 mm/day at Pusa. Among the ML models with radiation-based inputs, the SVR model performs the best at Pusa station, with RMSE of 0.345 mm/day. Considering the spatial transferability, the temperature-based CNN model developed for Pusa exhibited the highest transferability score of 1.0. The best performance by an ML model beyond the study station is observed for the radiation-based SVR model from Pusa: R2 = 0.895, d = 0.945, RMSE = 0.87 mm/day, MAPE = 0.150. These findings indicate that the application of ML models can significantly enhance the accuracy of ETO estimation while achieving effective transferability compared to empirical models. By addressing the limitations of traditional approaches, the methodology developed in this study can be applied in other regions to estimate ETO and to model other complex hydrological processes.