【摘要】 土壤含水率是影响农业生产和气候环境变化的重要因素。针对传统GNSS-R土壤含水率反演方法精度低、无法提取数据深度特征等局限，本文提出一种融合卷积神经网络(CNN)与长短期记忆网络(LSTM)的深度学习模型，用于土壤含水率反演。该模型可充分发挥CNN在特征提取与LSTM在时序建模方面的优势，尤其在多源数据融合方面的互补性。通过融合GNSS-R反射信号与微波辐射计数据(亮温(TB)、土壤粗糙度和植被覆盖指数(TR)),模型能更全面刻画土壤含水率变化趋势。与传统GNSS-R反演模型及多种机器学习方法(如支持向量机(SVM)、多层感知机(MLP)、随机森林(RF))相比，所构建模型相关系数(R)和均方根误差(RMSE)表现更优。试验验证结果表明，CNN-LSTM在精度、稳定性及泛化能力方面具有明显优势，适用于固定区域土壤含水率的连续监测，展示了深度神经网络在遥感数据融合与土壤含水率反演中的应用潜力。更多还原

【Abstract】 Soil moisture is a significant variable that influences agricultural productivity and climate dynamics. However, conventional GNSS-reflection(GNSS-R) methods for soil moisture retrieval often suffer from limited accuracy and insufficient capability to capture deep data features. A hybrid deep learning model that integrates convolutional neural networks(CNN) and long short-term memory(LSTM) networks for soil moisture retrieval was proposed to improve these issues. This model can fully leverage the advantages of CNN in feature extraction and LSTM in time series modeling, especially the complementarity in multi-source data fusion. By incorporating GNSS-R reflected signals along with brightness temperature(TB) and the integrated attenuation coefficient related to soil roughness and vegetation cover(TR) obtained from the ELBARA-Ⅱ radiometer, the proposed model enabled a more comprehensive representation of soil moisture dynamics. Compared with the conventional GNSS-R retrieval model and classical machine learning methods(such as support vector machine(SVM), multilayer perceptron(MLP), random forest(RF)), the proposed model achieved higher correlation coefficients(R) and lower root mean square errors(RMSE). The results verified that the proposed model offered superior accuracy, robustness, and generalization performance, making it well-suited for continuous soil moisture monitoring in fixed regions. This work demonstrated the potential of deep neural networks for enhancing remote sensing data fusion and improving the accuracy of soil moisture retrieval.更多还原