Scholarlink

<p>Accurate forecasting of root zone soil moisture (RZSM) is crucial for effective groundwater management, irrigation planning, and drought mitigation in semi-arid agrarian regions such as South Bihar. Traditional hydrological models mostly struggle to figure out the nonlinear temporal dynamics inherent in soil moisture data. This study proposes a hybrid deep ensemble learning framework that leverages the strengths of six deep neural networks (LSTM, Bi-LSTM, GRU, Bi-GRU, RNN, and CNN) as base models, with eXtreme Gradient Boosting (XGBoost) employed as a meta-learner in a stacked ensemble architecture. Each model is independently trained on Groundwater Root Zone Soil Wetness (GWETROOT) time series data (Jan 1985 to June 2025), and their predictions are aggregated using XGBoost to generate a robust final forecast. The study utilizes daily RZSM data (GWETROOT, surface to 100&#xa0;cm depth) obtained from the NASA POWER project, which provides satellite- and model-based gridded estimates. The performance of all models was evaluated across five districts in Bihar (Arwal, Aurangabad, Gaya, Jehanabad, and Nawada) using standard statistical metrics including MAE, MAPE, RMSE, MSE, and <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation>. Results demonstrate that the proposed ensemble approach consistently outperformed individual models, offering improved accuracy (<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(R^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> = 0.99) and generalizability. The findings highlight the effectiveness of integrating deep learning with ensemble techniques for soil moisture forecasting and offer a scalable solution for climate-resilient water resource management.</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 deep ensemble learning framework for forecasting root zone soil moisture in semi-arid regions of South Bihar

  • 期刊论文

  • 发表日期：

    2026年5月21日

查看全文
  • Ravi Patel,
  • Aditya Kumar,
  • Niharika Koch,
  • Gurpreet Singh,
  • Jainath Yadav

摘要

Accurate forecasting of root zone soil moisture (RZSM) is crucial for effective groundwater management, irrigation planning, and drought mitigation in semi-arid agrarian regions such as South Bihar. Traditional hydrological models mostly struggle to figure out the nonlinear temporal dynamics inherent in soil moisture data. This study proposes a hybrid deep ensemble learning framework that leverages the strengths of six deep neural networks (LSTM, Bi-LSTM, GRU, Bi-GRU, RNN, and CNN) as base models, with eXtreme Gradient Boosting (XGBoost) employed as a meta-learner in a stacked ensemble architecture. Each model is independently trained on Groundwater Root Zone Soil Wetness (GWETROOT) time series data (Jan 1985 to June 2025), and their predictions are aggregated using XGBoost to generate a robust final forecast. The study utilizes daily RZSM data (GWETROOT, surface to 100 cm depth) obtained from the NASA POWER project, which provides satellite- and model-based gridded estimates. The performance of all models was evaluated across five districts in Bihar (Arwal, Aurangabad, Gaya, Jehanabad, and Nawada) using standard statistical metrics including MAE, MAPE, RMSE, MSE, and \(R^2\) R 2 . Results demonstrate that the proposed ensemble approach consistently outperformed individual models, offering improved accuracy ( \(R^2\) R 2 = 0.99) and generalizability. The findings highlight the effectiveness of integrating deep learning with ensemble techniques for soil moisture forecasting and offer a scalable solution for climate-resilient water resource management.