Scholarlink

<p>Daily streamflow prediction is essential for effective water resources management, particularly in small, rapidly responding Mediterranean catchments characterised by pronounced seasonal variability and flashy hydrological regimes. This study evaluates two ensemble tree-based machine learning models; LightGBM and Random Forest, for daily discharge prediction in the Fiumarella di Corleto basin (33&#xa0;km²), Southern Italy, using 21 years of daily hydrometeorological observations (2002–2022). Predictor variables included lagged discharge, three-day moving averages, multi-day cumulative precipitation, and harmonic seasonal indicators derived from the day of year. Hyperparameters were optimised using an expanding walk-forward cross-validation approach to preserve the temporal structure of the time series. Model performance was primarily assessed using the Aras diagram, which quantifies total error and its decomposition into bias, variability, and correlation errors across both training and independent test periods; additional evaluation metrics included the Kling-Gupta Efficiency (KGE), Nash-Sutcliffe Efficiency (NSE), and Root Mean Squared Error (RMSE). LightGBM substantially outperformed Random Forest, achieving total errors of 5% in both periods according to the Aras diagram, compared to 21% for Random Forest. This superior performance was consistent with higher KGE (0.971 vs. 0.795), NSE (0.970 vs. 0.870), and lower RMSE (0.125 vs. 0.259&#xa0;mm day⁻¹), with near-zero systematic bias for both models (PBIAS &lt; 1%). SHAP analysis computed on the full test set identified the three-day moving average of antecedent discharge (Q_ma3) as the dominant predictor, followed by precipitation and lagged discharge variables, with cross-model agreement in feature rankings confirming the robustness of these findings. These results demonstrate that LightGBM, combined with rigorous temporal validation, Aras-based performance evaluation, and SHAP interpretability, provides a reliable and physically consistent framework for daily streamflow prediction in small Mediterranean basins.</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

Daily streamflow prediction using lightGBM and random forest in the Fiumarella di Corleto basin, Southern Italy

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月26日

查看全文
  • Marwah Yaseen,
  • Silvano Fortunato Dal Sasso,
  • Aras Izzaddin,
  • Maria Rosaria Margiotta,
  • Beniamino Onorati,
  • Biagio Sileo,
  • Vito Iacobellis

摘要

Daily streamflow prediction is essential for effective water resources management, particularly in small, rapidly responding Mediterranean catchments characterised by pronounced seasonal variability and flashy hydrological regimes. This study evaluates two ensemble tree-based machine learning models; LightGBM and Random Forest, for daily discharge prediction in the Fiumarella di Corleto basin (33 km²), Southern Italy, using 21 years of daily hydrometeorological observations (2002–2022). Predictor variables included lagged discharge, three-day moving averages, multi-day cumulative precipitation, and harmonic seasonal indicators derived from the day of year. Hyperparameters were optimised using an expanding walk-forward cross-validation approach to preserve the temporal structure of the time series. Model performance was primarily assessed using the Aras diagram, which quantifies total error and its decomposition into bias, variability, and correlation errors across both training and independent test periods; additional evaluation metrics included the Kling-Gupta Efficiency (KGE), Nash-Sutcliffe Efficiency (NSE), and Root Mean Squared Error (RMSE). LightGBM substantially outperformed Random Forest, achieving total errors of 5% in both periods according to the Aras diagram, compared to 21% for Random Forest. This superior performance was consistent with higher KGE (0.971 vs. 0.795), NSE (0.970 vs. 0.870), and lower RMSE (0.125 vs. 0.259 mm day⁻¹), with near-zero systematic bias for both models (PBIAS < 1%). SHAP analysis computed on the full test set identified the three-day moving average of antecedent discharge (Q_ma3) as the dominant predictor, followed by precipitation and lagged discharge variables, with cross-model agreement in feature rankings confirming the robustness of these findings. These results demonstrate that LightGBM, combined with rigorous temporal validation, Aras-based performance evaluation, and SHAP interpretability, provides a reliable and physically consistent framework for daily streamflow prediction in small Mediterranean basins.