Scholarlink

<p>Precise streamflow estimation in ungauged catchments remains critical for sustainable management of stream-wetland ecosystem. This study proposed a novel regionalization framework employing machine learning (ML) models that integrates Curve Number (CN) and specific discharge normalization, and compares it to a conventional SWAT parameter-transfer method. The framework is demonstrated on the Deepor Beel Ramsar wetland in Northeast India, whose contributing basins lack observed flow data. We designate the upstream Basistha basin as the donor catchment, and developed and systematically evaluated five ML models; Support Vector Regression (SVR), Extreme Gradient Boosting (XGB), Random Forest (RF), LightGBM (LGBM), and Artificial Neural Networks (ANN) utilizing recursive feature elimination with cross‑validation (RFECV) to select optimal model-specific predictor subsets. RFECV result indicates that SVR exhibit strong performance with four essential hydrometeorological variables, while tree-based methods consistently incorporated CN alongside precipitation and temperature, underscoring CN’s vital role in capturing catchment infiltration dynamics. Sensitivity analysis of six CN estimation scenarios indicated that linear interpolation from 1992 to 2002 yields the best RMSE, bias, NSE and R². In the donor basin, LGBM surpassed both alternative ML models and the calibrated SWAT model. Upon regionalization to the target basin through specific discharge scaling, the ML-CN methodology produced R² = 0.804, NSE = 0.804, and PBIAS = -2.73%, outperforming the SWAT transfer results (R² = 0.708; NSE = 0.613; PBIAS = + 9.44%). Integrating robust predictor selection, specific‑discharge normalization, and dynamic CN within an ML framework will provide a straightforward yet powerful tool for accurate streamflow prediction in ungauged, data‑scarce basin.</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

An Easy-to-Apply Machine Learning Framework for Hydrologic Evaluation of Ungauged Catchments

  • 期刊论文

  • 发表日期：

    2025年12月29日

查看全文
  • Bhaswatee Baishya,
  • Arup Kumar Sarma

摘要

Precise streamflow estimation in ungauged catchments remains critical for sustainable management of stream-wetland ecosystem. This study proposed a novel regionalization framework employing machine learning (ML) models that integrates Curve Number (CN) and specific discharge normalization, and compares it to a conventional SWAT parameter-transfer method. The framework is demonstrated on the Deepor Beel Ramsar wetland in Northeast India, whose contributing basins lack observed flow data. We designate the upstream Basistha basin as the donor catchment, and developed and systematically evaluated five ML models; Support Vector Regression (SVR), Extreme Gradient Boosting (XGB), Random Forest (RF), LightGBM (LGBM), and Artificial Neural Networks (ANN) utilizing recursive feature elimination with cross‑validation (RFECV) to select optimal model-specific predictor subsets. RFECV result indicates that SVR exhibit strong performance with four essential hydrometeorological variables, while tree-based methods consistently incorporated CN alongside precipitation and temperature, underscoring CN’s vital role in capturing catchment infiltration dynamics. Sensitivity analysis of six CN estimation scenarios indicated that linear interpolation from 1992 to 2002 yields the best RMSE, bias, NSE and R². In the donor basin, LGBM surpassed both alternative ML models and the calibrated SWAT model. Upon regionalization to the target basin through specific discharge scaling, the ML-CN methodology produced R² = 0.804, NSE = 0.804, and PBIAS = -2.73%, outperforming the SWAT transfer results (R² = 0.708; NSE = 0.613; PBIAS = + 9.44%). Integrating robust predictor selection, specific‑discharge normalization, and dynamic CN within an ML framework will provide a straightforward yet powerful tool for accurate streamflow prediction in ungauged, data‑scarce basin.