Scholarlink

<p>Accurate prediction of soil moisture (SM) and evapotranspiration (ET) is essential for water resource management. While Long Short-Term Memory (LSTM) networks are widely applied, they struggle with long-term temporal dependencies and predictive uncertainty. To address these limitations, we developed the Multi-Task Learning and Multi-Expert Gating (MTMEG) module, which employs a Transformer to capture long-range dynamics and leverages multi-task learning to mitigate predictive uncertainty by coupling SM and ET.Comprehensive evaluations of 1- and 3-day forecasts demonstrate that MTMEG significantly outperforms the baseline LSTM. For 1-day lead times, R² improved by 5.13% (SM) and 6.83% (ET), with Kling-Gupta Efficiency (KGE) increasing by 6.77% and 5.53%, respectively. Moreover, the integration of the MTMEG module into various advanced LSTM architectures corroborated its broad applicability. These findings underscore the proposed framework as a highly adaptable solution that simultaneously enhances predictive accuracy and robustness, while enabling seamless integration across diverse modeling architectures.</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

Improving Short-term Forecasts of Soil Moisture and Evapotranspiration with Multi-task Learning and Transformer-based Feature Processing

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Yuguang Yan,
  • Gan Li,
  • Qingliang Li,
  • Xiao Chen,
  • Jinlong Zhu

摘要

Accurate prediction of soil moisture (SM) and evapotranspiration (ET) is essential for water resource management. While Long Short-Term Memory (LSTM) networks are widely applied, they struggle with long-term temporal dependencies and predictive uncertainty. To address these limitations, we developed the Multi-Task Learning and Multi-Expert Gating (MTMEG) module, which employs a Transformer to capture long-range dynamics and leverages multi-task learning to mitigate predictive uncertainty by coupling SM and ET.Comprehensive evaluations of 1- and 3-day forecasts demonstrate that MTMEG significantly outperforms the baseline LSTM. For 1-day lead times, R² improved by 5.13% (SM) and 6.83% (ET), with Kling-Gupta Efficiency (KGE) increasing by 6.77% and 5.53%, respectively. Moreover, the integration of the MTMEG module into various advanced LSTM architectures corroborated its broad applicability. These findings underscore the proposed framework as a highly adaptable solution that simultaneously enhances predictive accuracy and robustness, while enabling seamless integration across diverse modeling architectures.