Scholarlink

<p>Spatiotemporal forecasting of surface soil moisture (SSM) is recognized as a critical scientific issue in precision agricultural irrigation, regional drought monitoring, and early warning systems for extreme precipitation. However, long-term forecasting continues to pose formidable challenges because of the complexity observed across both the spatial and temporal scales. In this study, we used a daily SSM dataset at a 0.05°×0.05° spatial resolution over the Qilian Mountains, China and proposed a hybrid Convolutional Long Short-Term Memory (ConvLSTM)-Nudging model, which combined deep neural networks with data assimilation to increase the accuracy of long-term SSM forecasting. We trained and evaluated the SSM predictive performance of four models (Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), ConvLSTM, and ConvLSTM with Squeeze-and-Excitation (SE) attention mechanism (ConvLSTM-SE)) in both short-term and long-term scenarios. The results showed that all the models perform well under short-term predictions, but the accuracy decrease substantially in long-term predictions. Therefore, we integrated Nudging technique during the long-term prediction phase to assimilate observational information and rectify model biases. Comprehensive evaluations demonstrate that Nudging significantly improves all the models, with ConvLSTM-Nudging achieving the best performance under the 200-d forecasting scenario. Relative to those of the best-performing ConvLSTM model for long-term forecasts, when observation noise <i>δ</i>=0.00 and observation fraction <i>obs</i>=50.0%, the coefficient of determination (<i>R</i><sup>2</sup>) of ConvLSTM-Nudging increases by approximately 82.1%, while its mean absolute error (MAE) and root mean squared error (RMSE) decrease by approximately 84.8% and 77.3%, respectively; the average Pearson correlation coefficient (<i>r</i>) improves by approximately 23.6%, and Bias is reduced by 98.1%. These results demonstrated that although pure deep learning models achieve high accuracy in the short-term predictions, they are prone to error accumulation and systematic drift in long-term autoregressive predictions. Integrating data assimilation with deep learning and continuously correcting the state through observation can effectively suppress long-term biases, thereby achieving robust long-term SSM forecasting.</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 hybrid ConvLSTM-Nudging model for predicting surface soil moisture in the Qilian Mountains, China

  • 期刊论文

  • 发表日期：

    2025年11月20日

查看全文
  • Manhong Fan,
  • Qian Xiao,
  • Qinghe Yu,
  • Junhao Zhao

摘要

Spatiotemporal forecasting of surface soil moisture (SSM) is recognized as a critical scientific issue in precision agricultural irrigation, regional drought monitoring, and early warning systems for extreme precipitation. However, long-term forecasting continues to pose formidable challenges because of the complexity observed across both the spatial and temporal scales. In this study, we used a daily SSM dataset at a 0.05°×0.05° spatial resolution over the Qilian Mountains, China and proposed a hybrid Convolutional Long Short-Term Memory (ConvLSTM)-Nudging model, which combined deep neural networks with data assimilation to increase the accuracy of long-term SSM forecasting. We trained and evaluated the SSM predictive performance of four models (Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), ConvLSTM, and ConvLSTM with Squeeze-and-Excitation (SE) attention mechanism (ConvLSTM-SE)) in both short-term and long-term scenarios. The results showed that all the models perform well under short-term predictions, but the accuracy decrease substantially in long-term predictions. Therefore, we integrated Nudging technique during the long-term prediction phase to assimilate observational information and rectify model biases. Comprehensive evaluations demonstrate that Nudging significantly improves all the models, with ConvLSTM-Nudging achieving the best performance under the 200-d forecasting scenario. Relative to those of the best-performing ConvLSTM model for long-term forecasts, when observation noise δ=0.00 and observation fraction obs=50.0%, the coefficient of determination (R2) of ConvLSTM-Nudging increases by approximately 82.1%, while its mean absolute error (MAE) and root mean squared error (RMSE) decrease by approximately 84.8% and 77.3%, respectively; the average Pearson correlation coefficient (r) improves by approximately 23.6%, and Bias is reduced by 98.1%. These results demonstrated that although pure deep learning models achieve high accuracy in the short-term predictions, they are prone to error accumulation and systematic drift in long-term autoregressive predictions. Integrating data assimilation with deep learning and continuously correcting the state through observation can effectively suppress long-term biases, thereby achieving robust long-term SSM forecasting.