作为全球基本气候变量之一，土壤湿度在地球科学研究和全球气候变化监测中发挥着至关重要的作用，不仅影响着水文循环、径流形成、土壤侵蚀、作物产量预测和许多其他进程，还能够通过大气反馈来影响地表与大气之间水和能量的相互交换，同时土壤湿度也是全球陆地蒸散的一个重要水源。因此高时空分辨率的表层土壤湿度信息对于农业干旱监测、灌溉管理和数值天气预报等领域的实际应用和科学研究具有重要意义。然而，目前基于地面站点、微波遥感和模型同化技术的土壤湿度产品受固有空间分辨率和产品精度的限制，难以满足精细领域的应用需求。因此，本文旨在探索利用多源数据融合和机器学习方法进行土壤湿度的降尺度研究。本文的具体研究内容如下：

（1）基于多源数据融合的土壤湿度降尺度

与单一数据源相比，多源数据融合能够有效提升产品的时空分辨率，充分发挥不同数据源的独特优势。因此，本文利用基于球冠谐分析和赫尔模特方差分量估计的融合方法集成多种数据生成了空间分辨率为1 km的降尺度产品。该方法首先利用球冠谐方程将多种数据拟合至球冠表面，构建出关于未知参数的观测方程。然后使用HVCE方法确定不同类型数据的相对权重，并通过最小二乘法求解未知参数。最后通过融合多种土壤湿度数据从而获得高质量的降尺度产品。之后，本文分别利用地面站点数据和模型产品对降尺度产品进行了精度评估。结果表明，在不同的气候条件下，降尺度产品与站点数据之间的相关系数始终高于其他产品，性能表现是最好的。在与验证站的对比中，降尺度产品的精度显著高于其它产品。在空间特征对比中，降尺度产品能够准确捕获地面站点的土壤湿度的空间变化，具有更丰富的细节特征。

（2）基于机器学习的土壤湿度降尺度

以往的研究表明，土壤湿度的时空变化易受多种外界因素的影响，具有复杂的空间异质性。因此，本文共选取12种与土壤湿度相关的特征参量，采用四种不同的基于机器学习的降尺度方法（随机森林（RF）、反向传播神经网络（BPNN）、残差神经网络（ResNet）和混合模型（CNN-GRU））将SMAP土壤湿度产品的空间分辨率从原始的36 km降至1 km，并利用站点数据和其他产品对降尺度产品进行了评估分析。结果表明，四种降尺度产品均能够与原始SMAP产品保持高度的空间一致性，还提供了更详细的空间信息。与站点数据相比，RF降尺度产品在总体上具有较高的相关性和较低的偏差，是所有产品中表现最好的。就不确定性分析结果来看，除BPNN降尺度产品外，其他降尺度产品均具有较低的不确定性，这意味着它们能够提供更可靠的土壤湿度信息。此外，进一步的评估发现，与其他产品相比，RF降尺度产品能够更有效捕捉土壤湿度的时空变化，且与降水数据表现出较高的相关性。

As one of the fundamental global climate variables, soil moisture plays a crucial role in earth science research and global climate change monitoring. It not only affects hydrological cycles, runoff formation, soil erosion, crop yield prediction, and many other processes, but also affects the exchange of water and energy between the surface and the atmosphere through atmospheric feedback. At the same time, soil moisture is also an important water source for global land evapotranspiration. Therefore, high spatiotemporal resolution information on surface soil moisture is of great significance for practical applications and scientific research in fields such as agricultural drought monitoring, irrigation management, and numerical weather forecasting. However, current soil moisture products based on ground stations, microwave remote sensing and model assimilation technologies are limited by inherent spatial resolution and product accuracy, making them difficult to meet the application needs of fine fields. Hence, this article aims to explore downscaling research on soil moisture using multi-source data fusion and machine learning methods. The specific research contents of this study are as follows:

(1) Soil Moisture Downscaling Based on Multi-source Data Fusion

Compared with single data source, multi-source data fusion can effectively improve the spatiotemporal resolution of products and fully leverage the unique advantages of different products. Therefore, this article integrates multiple data using a fusion method based on Spherical Cap Harmonic Analysis (SCHA) and Helmert Variance Component Estimation (HVCE) to generate downscaling products with a spatial resolution of 1 km. This method first utilizes the spherical cap harmonic equation to fit various data to the surface of the spherical cap, and constructs observation equations about unknown parameters. Then, the HVCE method is used to determine the relative weights of different types of data and solve unknown parameters through least squares. Finally, high-quality soil moisture fusion products are obtained by integrating multiple soil moisture data. Afterwards, this article evaluated the accuracy of the downscaling products using ground station data and model products. The results indicate that under different climate conditions, the correlation coefficient between the downscaling products and the station data is consistently higher than other products, and the performance is the best. In comparison with the validation stations, the accuracy of the downscaling products is significantly better than other products. In spatial feature comparison, the downscaling products can accurately capture the spatial changes in soil moisture at ground sites, with richer detailed features.

(2) Soil Moisture Downscaling Based on Machine Learning

Previous studies have shown that the spatiotemporal variations of soil moisture are susceptible to various external factors and exhibit complex spatial heterogeneity. Hence, this article selected a total of 12 feature parameters related to soil moisture and adopted four different machine learning-based downscaling methods (Random Forest (RF), Back Propagation Neural Network (BPNN), Residual Neural Network (ResNet), and Hybrid Model (CNN-GRU)) to downscale the spatial resolution of SMAP soil moisture products from the original 36 km to 1 km. Then, the downscaled products were evaluated and analyzed through site data and other products. The results show that all four downscaled products can maintain high spatial consistency with the original SMAP products and provide more detailed spatial information. Compared to site data, the RF downscaled product shows higher overall correlation and lower bias, performing the best among all products. In terms of uncertainty analysis, except for the BPNN downscaled product, the other downscaled products have lower uncertainty, indicating that they can provide more reliable soil moisture information. Furthermore, further evaluation found that compared to other products, the RF downscaled products can more effectively capture the spatiotemporal variations of soil moisture and exhibit higher correlation with precipitation data.