Scholarlink

<p>Latent heat flux (LE) plays a crucial role in the global water cycle and surface energy exchange. The accuracy of machine learning models depends on the quality of input data. Therefore, this study proposes a physical model-assisted ensemble learning (EML) framework to improve data quality during the training process. The Heihe River Basin (HRB) in Northwest China was selected as the study area. The joint models were compared with the physical model (TSEB) and three EML models (XGBoost, LightGBM, CatBoost). The SHAP method was used to interpret the results of the joint models and quantify the contributions of input features to LE estimation. The modeling errors of different sites in the joint model were evaluated and retrieved in the region. The results show that EML models exhibited good fitting and generalization capabilities, with good accuracy in terms of MAE and RMSE. The XGBoost model showed an improvement in estimation accuracy by 4.20–8.92% and 7.11–13.88% compared to LightGBM and CatBoost models, respectively. The joint models improved accuracy by 0.2–4.74% over the individual EML models, with the TSEB-assisted XGBoost model demonstrating the best overall performance. The joint models effectively captured the impacts of energy, temperature, moisture, and vegetation on LE in the HRB. The joint models were within an acceptable range when modeling at all sites, and the regional spatial patterns of LE were consistent.</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

A dual-source energy balance model-assisted ensemble learning approach for estimating latent heat flux

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Lizheng Wang,
  • Jinling Kong,
  • Qiutong Zhang,
  • Yanling Zhong

摘要

Latent heat flux (LE) plays a crucial role in the global water cycle and surface energy exchange. The accuracy of machine learning models depends on the quality of input data. Therefore, this study proposes a physical model-assisted ensemble learning (EML) framework to improve data quality during the training process. The Heihe River Basin (HRB) in Northwest China was selected as the study area. The joint models were compared with the physical model (TSEB) and three EML models (XGBoost, LightGBM, CatBoost). The SHAP method was used to interpret the results of the joint models and quantify the contributions of input features to LE estimation. The modeling errors of different sites in the joint model were evaluated and retrieved in the region. The results show that EML models exhibited good fitting and generalization capabilities, with good accuracy in terms of MAE and RMSE. The XGBoost model showed an improvement in estimation accuracy by 4.20–8.92% and 7.11–13.88% compared to LightGBM and CatBoost models, respectively. The joint models improved accuracy by 0.2–4.74% over the individual EML models, with the TSEB-assisted XGBoost model demonstrating the best overall performance. The joint models effectively captured the impacts of energy, temperature, moisture, and vegetation on LE in the HRB. The joint models were within an acceptable range when modeling at all sites, and the regional spatial patterns of LE were consistent.