土壤水分是全球水循环和地-气能量交换的重要参数之一，是水文、气象、农业、生态等环境研究中的关键因子。目前被动微波遥感空间分辨率较低（25km~40km），仅适用于全球尺度研究，无法满足干旱监测、生态系统建模、水文气候模型分析（＜10km）等区域尺度上的实际需求。因此，如何提高土壤水分空间分辨率是本文研究的重点。本文针对农林混合区，发展高空间分辨率的土壤水分反演算法，以期提高获取高空间分辨率的土壤水分。研究内容主要包括：

（1）Sentinel-1和ALOS-2雷达土壤水分反演算法：基于参数优化后的水云模型和AIEM模型，Oh模型分别构建了针对C波段和L波段的地表综合模拟数据库。在参数敏感性分析的基础上，选择了模拟数据库中的同极化后向散射系数（VV或HH）、入射角、NDVI和土壤水分数据利用神经网络进行了训练和测试。基于神经网络训练的最优结果，分别利用ALOS-2和Sentinel-1 SAR的后向散射、入射角和NDVI反演获得了空间分辨率30m的ALOS-2和Sentinel-1土壤水分。结果表明，Sentinel-1和ALOS-2土壤水分在低植被地表有较高精度。ALOS-2土壤水分在高植被区的反演精度高于Sentinel-1土壤水分。ALOS-2 L波段SAR在农林交错地区土壤水分的反演中比Sentinel-1 C波段SAR更具有潜力。

（2）光学/热红外遥感与被动微波遥感结合的土壤水分降尺度算法：该算法针对光学/热红外数据地表温度缺失情况，引入了时空连续的地表温度数据TRIMS LST；针对光学数据NDVI饱和点比LAI低的问题，使用GLASS LAI代替了NDVI；为了加强地表参数与土壤水分的关系，克服光学数据的易饱和问题，引入了微波极化差指数MPDI，基于多元回归构建了SMAP L3被动微波土壤水分降尺度方法，获取了1km空间分辨率的土壤水分。单站点验证结果表明，降尺度的土壤水分在草地区的精度高于灌木和林地。草地区，降尺度土壤水分与实测土壤水分的相关性为0.6，RMSE为0.055m³m^-3。灌木和林区验证中，相关系数分别为0.560，0.520，RMSE分别为0.069 m³m^-3，0.072 m³m^-3。整体上，降尺度后的土壤水分在空间分布上体现了更多的细节信息。

（3）主被动微波遥感结合的土壤水分降尺度算法：该方法基于地表综合模拟数据库，利用后向散射系数、后向散射极化比值与土壤水分的关系，构建了SMAP L3被动微波土壤水分降尺度方法，获取了1km空间分辨率的土壤水分。单站点验证表明，降尺度后的土壤水分数据和实测土壤水分具有较高的相关性，且草地和灌木区的土壤水分降尺度精度高于林地，在草地地表的相关系数为0.691，RMSE为 0.52 m³m^-3；在灌木和林地的相关系数分别为0.606，0.561，RMSE分别为0.63 m³m^-3，0.67 m³m^-3。基于地面站点实测数据和SMAP L2 SP土壤水分数据，对上述光学热红外遥感与被动微波遥感结合的土壤水分降尺度算法和主被动微波遥感结合的土壤水分降尺度算法进行了比较和分析，结果表明，主被动微波遥感结合的方法在根河地区1km空间分辨率的土壤水分估算中比光学/热遥感和被动微波数据结合的方法更具有潜力。

Soil moisture content (SMC) is one of the important parameters of global water cycle and land-atmosphere energy exchange, and it is a key parameter in hydrology, meteorology, agriculture, ecology, and other environmental studies. At present, passive microwave remote sensing have a low spatial resolution (25km~40km), which is applicable to global scale research and cannot meet the actual needs of regional standards（＜10km）such as drought monitoring, ecosystem modeling, and hydro-climatic model analysis. It is difficult to meet application requirements on the regional scale. Therefore, the focus of this paper is on how to improve the spatial resolution of SMC. In this paper, SMC retrieval methods with high spatial resolution are developed in agroforestry areas. The contents of this paper mainly include the following parts:

  (1) Sentinel-1 and ALOS-2 SAR SMC retrieval algorithm. The water cloud model with optimized parameters, AIEM model, and Oh model was used to construct the land surface synthetic database for C-band and L-band, respectively. According to the sensitivity analysis of parameters, the co-polarization backscattering coefficient (VV or HH), incident Angle, NDVI and SMC in the simulation database were selected for training and testing using the artificial neural network (ANN). Based on the optimal results of neural network training, the SAR co-polarization backscatter, the local incidence angle, and the NDVI were used in input vectors for the ANN algorithm for the retrieval and mapping of the ALOS-2 and Sentinel-1 SMC at a 30 m resolution.

The results showed that the accuracy of Sentinel-1 and ALOS-2 SMC is high in low vegetation area. The accuracy of ALOS-2 SMC is higher than that of Sentinel-1 SMC in high vegetation area. ALOS-2 L-band SAR appears to have more potential than the Sentinel-1 C-band SAR to estimate soi moisture in agroforestry areas.

  (2) Fused passive microwave data and optical/thermal data to downscale SMAP L3 SMC product. Aiming at the space-time missing problem caused by the influence of weather on thermal infrared data, the space-time continuum LST data was used in this method. Because of the saturation point of optical data NDVI is lower than LAI, we use the GLASS LAI instead of NDVI. At the same time, we introduce the MPDI to overcome the problem of saturation of optical data. And the SMAP L3 passive microwave SMC downscaling algorithm was constructed based on polynomial regression method to obtain SMC with a spatial resolution of 1km. The results show that the accuracy of downscaled SMC in grassland is higher than that of shrubland and forestland. The R and RMSE between downscaled SMC and in situ SMC in grassland are 0.6, 0.055m³m^-3, respectively. For the shrub and forestland, the R are 0.560, 0.520, and RMSE are 0.069m³m^-3, 0.072m³m^-3, respectively. Moreover, the downscaled SMC can present more details in the spatial distribution.

  (3) Fused active and passive microwave data to downscale SMAP L3 SMC product. Based on the land surface synthetic database, the SMAP L3 passive microwave SMC downscaling algorithm was constructed by the relationship between the backscattering coefficient, the ratio of backscattering polarization and SMC to obtain SMC with a spatial resolution of 1km. The results show that the accuracy of downscaled SMC in grassland is higher than that of shrubland and forestland. The R and RMSE between downscaled SMC and in situ SMC in grassland are 0.691, 0.052m³m^-3, respectively. For the shrub and forestland, the R are 0.606，0.561, and RMSE are 0.63 m³m^-3，0.67 m³m^-3, respectively. Two SMC downscaling methods were evaluated by the in-situ measurements and compared by the SMAP L2 SP SMC product in agroforestry areas. The results show that the active and passive microwave combination method has more potential than passive microwave and optical/thermal remote sensing combination method to produce high-resolution SMC in Genhe area.