联合机器学习和多尺度集合卡尔曼滤波算法的积雪数据同化方法研究

The objective of this project is to exploring the potential of applying machine learning methods (neural network, support vector machine, and deep learning) to the study of snow data assimilation. The northern Xinjiang is selected as study area, which is a typical seasonal snow area in China. Firstly, we will develop "snow cover area - snow depth" depletion curve model based on machine learning methods to reduce the uncertainty of the observation operator in snow cover area data assimilation. Then, we will develop the agent model of multi-layer snow microwave radiative transfer model (MEMLS) based on machine learning methods to simplify model structure, reduce the input variables required by the model, and improve computation efficiency. Finally, the multiscale ensemble Kalman filter algorithm will be used to realize the joint assimilation of snow cover area and passive microwave bright temperature. The project will help to promote and expand the application of multi-source remote sensing data in the field of hydrology, and provide an important method for improving regional prediction of snow depth.

以我国典型的季节性积雪区域—北疆为研究区，重点探索如何将机器学习方法（神经网络、支持向量机及深度学习）应用到积雪数据同化研究中。我们将发展基于机器学习方法的“积雪面积-雪深”衰减曲线模型，减小积雪面积同化中观测算子的不确定性；发展基于机器学习方法的多层积雪微波辐射传输模型的代理模型，简化模型结构，减少模型所需的输入变量，提高计算效率；最后采用多尺度集合卡尔曼滤波算法实现积雪面积和被动微波亮温的联合同化。本项目的开展，有助于促进和拓展多源遥感数据在水文领域的深入应用，为改进区域雪深估算提供重要方法。

本项目以我国典型的季节性积雪区域—北疆为研究区，重点探索了如何将机器学习方法（神经网络、支持向量机及深度学习）应用到积雪参量反演与积雪数据同化研究中，在积雪面积产品重建、积雪面积数据同化、积雪模型多参数化方案评价方面取得了重要成果。具体包括：（1）发展了基于机器学习的积雪面积比例产品重建、积雪产品去云、以及雪深反演方法；（2）构建了基于机器学习的“积雪面积-雪深”关系曲线，并利用北疆观测数据进行了验证；（3）评价了Noah-MP模型积雪参数方案的敏感性与不确定性；（4）构建了多源遥感水文数据同化系统和基于“积雪面积-雪深”关系曲线的积雪面积同化方法。在项目执行期间，共发表论文10篇（SCI论文8篇，中文核心论文2篇），培养博士1名、硕士1名。