土壤水分作为陆地表面水资源形成、转化和消耗过程中的基本参数，也是地表能量交换的基本要素。遥感技术可以及时准确地监测区域尺度土壤水分，为社会生产管理提供有力保障。一方面，土壤水分是制约半干旱区植被生长和恢复的重要因素之一，另一方面，土壤水分也是主导农作物生长发育的重要因素之一。不同土壤水分模型反演结果受地域和时相选择等因素影响较大，精度不够稳定。因此，探讨适合不同气候类型不同地域的客观、实时和动态的土壤水分监测模型与反演方法，为社会经济发展和生产管理提供强有力的决策依据，具有重要的科学意义和现实意义。本研究主要利用四种特征空间的双抛物线型TVDI（Temperature vegetation dryness index）、四种光谱特征空间法（SMMI-Soil moisture monitoring index、PDI-Perpendicular drought index、MSMMI-Modified soil moisture monitoring index和MPDI-Modified perpendicular drought index）、一种引入组合粗糙度（Rs）的半经验SAR模型及四种机器学习回归模型（GRNN-Generalized neural network regression、RFR-Random forest regression、SVR-Support vector regression和DNNR-Deep neural network regression）监测地表土壤水分，并在两个研究区进行实验与应用，一是中国陕西省神东矿区（温带大陆性半干旱区），二是加拿大曼尼托巴省农业区（温带大陆性湿润区），主要研究内容及结果如下：
（1）主要分析了四种双抛物线型TVDI监测温带大陆性半干旱区土壤水分的性能，并利用被动微波遥感土壤水分数据进行交叉验证，结果表明，Ts（Land surface temperature）-NDVI（Normalized difference vegetation index）/Ts-EVI（Enhanced vegetation index）特征空间监测半干旱区土壤水分比Ts-FPAR（Fraction of photosynthetically active radiation）/Ts-LAI（Leaf area index）特征空间更具稳健性。降雨量是影响该区土壤水分变化及植被生长的限制性因素。21世纪以来，神东矿区的植被恢复效果显著，植被覆盖度呈稳步上升的趋势，温室气体对植被恢复有一定的促进作用。2000~2011年神东矿区地表土壤水分有所增高，2012年达到峰值，体积土壤水分百分比年均值为28.87%，近五年地表土壤水分有所减小，逐渐恢复到21世纪初的水平，受极端气候影响较明显。
进一步探讨了四种双抛物线型TVDI监测温带大陆性湿润区土壤水分的性能，并利用原位土壤水分和土壤温度等数据进行验证，结果表明，对于监测该区作物生长前期的土壤水分，Ts-NDVI特征空间的效果更好。Ts-EVI/Ts-FPAR特征空间更适合监测该区作物生长中/后期的土壤水分，且两者监测土壤水分的稳健性较高。TVDI对土壤温度比土壤水分更加敏感，影响该区土壤水分变化的主导因素主要为降雨。
（2）主要分析了光谱特征空间法（SMMI、PDI、MSMMI和MPDI）监测温带大陆性半干旱区土壤水分的性能，结果表明，与实测土壤水分相比，这四个指数均能有效反演植被覆盖度较低的半干旱区地表浅层土壤水分，且引入短波红外的特征空间反演精度更高。尺度化MSMMI与双抛物线型TVDI变化趋势基本一致，与降雨量显著相关。
进一步分析了光谱特征空间法监测温带大陆性湿润区土壤水分的性能，并评估了植被覆盖度反演方式对MSMMI和MPDI监测土壤水分的影响。结果表明，这四个指数均能有效反演该区0~5cm深裸土土壤水分，且引入短波红外的特征空间反演精度更高。SMMI和PDI没有考虑植被覆盖的影响，监测植被覆盖区土壤水分的准确性略低。引入红边、短波红外波段和植被覆盖度的MSMMI是监测该区土壤水分的最佳土壤水分指数，无论植被覆盖区、裸土区还是植被裸土混合区，其与实测0~5cm深土壤水分的相关性都是最高的，R分别为0.732、0.865和0.780。
（3）主要分析了滤波方式对Sentinel-1A双极化SAR后向散射系数反演像元尺度（20m空间分辨率）土壤水分的影响，并对比半经验SAR模型（引入组合粗糙度Rs）、光谱特征空间法（MSMMI和MPDI）以及机器学习回归模型反演温带大陆性湿润区土壤水分的性能与异同。结果表明，Lee sigma滤波法可以作为Sentinel-1双极化后向散射系数反演土壤水分的最佳滤波方式。光谱特征空间法反演高植被覆盖区土壤水分时存在明显的低估现象，但在反演裸土区土壤水分时，精度略高于半经验SAR模型。SAR数据更适合反演植被覆盖区土壤水分，在高植被覆盖区尤为明显。联合Sentinel-1和Sentinel-2反演土壤水分时，土壤水分反演值与实测值的R为0.8886，RMSE为0.0391cm³/cm³，明显优于单独使用光学或SAR数据。机器学习回归模型反演土壤水分的精度明显高于半经验SAR模型和光谱特征空间法，GRNN和RFR算法在反演精度和效率上优于SVR算法。结合传统四波段光谱特征（蓝、绿、红和近红外）和双极化SAR特征（VV、VH和雷达入射角）就可以准确地反演土壤水分，且基于RFR所得反演值与实测值的R可达0.9355，RMSE可达0.0284cm³/cm³。利用DNNR算法结合光学和SAR数据反演土壤水分时，RMSE可达0.0045cm³/cm³。联合Sentinel-1和Sentinel-2数据反演区域尺度10m~20m空间分辨率土壤水分具有很大的潜力。
 

~Soil moisture (SM), as a basic parameter in the process of formation, transformation and consumption of land surface water resources, is also the basic element of surface energy exchange. Remote sensing technology can timely and accurately monitor SM at regional scale, which provides strong guarantees for social production management. On the one hand, SM is one of the important factors restricting the growth and recovery of vegetation in semi-arid region. On the other hand, SM is also one of the important factors leading the growth and development of crops. Estimated results of different SM models are greatly affected by regional and temporal selection and the accuracy is not stable. Therefore, it is of great scientific and practical significance to explore objective, real-time and dynamic SM monitoring models and estimation methods suitable for different climate types and different regions, so as to provide a strong decision-making basis for social and economic development and production management. In this study, the double-parabolic TVDI (Temperature vegetation dryness index) of four feature spaces, four spectral feature space methods (SMMI-Soil moisture monitoring index, PDI-Perpendicular drought index, MSMMI-Modified soil moisture monitoring index, MPDI-Modified perpendicular drought index), a semi-empirical SAR model with combined roughness (Rs), and four machine learning regression models (GRNN-Generalized neural network regression, RFR-Random forest regression, SVR-Support vector regression, DNNR-Deep neural network regression) were employed to monitor surface SM in two study areas. One is Shendong mining area in Shaanxi province, China (Temperate continental semi-arid zone). The other is Manitoba agricultural area in Canada (Temperate continental humid zone). The main research contents and results are as follows:
(1) In this study, the performance of four dual-parabolic TVDI for monitoring SM in temperate continental semi-arid zone was analyzed, and cross-validation was carried out by using passive microwave remote sensing SM data. The results show that, Ts-NDVI/Ts-EVI feature space are more robust than Ts-FPAR/Ts-LAI feature space in monitoring SM in this area. Precipitation is a limiting factor that affects SM variation and vegetation growth in this area. Since the 21st century, the effect of vegetation restoration in Shendong mining area is remarkable, and the vegetation coverage shows a trend of steady rise. Greenhouse gases have a certain promotion effect on vegetation restoration. The surface SM in Shendong mining area increases from 2000 to 2011, and reaches the peak value in 2012. The annual mean value of volume SM percentage is 28.87%. The surface SM decreases in recent five years and gradually recoveres to the level of the early 21st century. The area is obviously affected by extreme climate.
Furthermore, the performance of four double-parabolic TVDI for monitoring SM in temperate continental humid zone was analyzed, and validation was conducted with in-situ SM and soil temperature data. The results show that, Ts-NDVI feature space is more effective for monitoring SM in the early stage of crop growth in this area. Ts-EVI/Ts-FPAR feature space are more suitable for monitoring SM in the middle and later growth period of crops, and both of them have higher robustness in monitoring SM. TVDI is more sensitive to soil temperature than SM, and precipitation is the main factor affecting SM in this area.
(2) The performance of the spectral feature space methods (SMMI, PDI, MSMMI MPDI) for monitoring SM in temperate continental semi-arid zone was analyzed. The results show that, compared with the measured SM, the four indices can effectively estimate surface SM in semi-arid zone with low vegetation coverage. The estimation accuracy of feature space with short-wave infrared is higher. The variation trend of scale-MSMMI was basically the same as that of the double parabolic TVDI, and it is significantly correlated with precipitation.
Furthermore, the performance of the spectral feature space methods (SMMI, PDI, MSMMI MPDI) for monitoring SM in temperate continental humid zone was analyzed. The effects of FVC (Fractional vegetation cover) estimation methods on estimating SM by MSMMI and MPDI were evaluated. The results show that, the four indices can be used to effectively estimate SM in bare soil areas at a depth of 0~5cm, and the feature space with short-wave infrared is still more accurate. The accuracy of monitoring SM in vegetation-covered areas was slightly lower by SMMI and PDI which do not consider the effect of vegetation cover. MSMMI with red edge, short-wave infrared bands and FVC is the best soil moisture index for monitoring SM in this area. The correlation between MSMMI and measured SM at a depth of 0~5cm is the highest in vegetation-covered areas, bare soil areas or vegetation-bare soil mixed areas, with the R values of 0.732, 0.865 and 0.780, respectively.
(3) This study mainly analyzed the influence of filtering methods on estimating SM at pixel scale (20m spatial resolution) based on the backscattering coefficients of Sentinel-1 dual-polarization SAR data. Then, the performance of the semi-empirical SAR model with combined roughness (Rs), spectral feature space methods (MSMMI and MPDI) and machine learning regression models for monitoring SM in temperate continental humid zone was analyzed. The results show that, Lee sigma filter is the best filtering method for estimating SM by Sentinel-1 dual-polarization backscattering coefficients. There is an obvious underestimation of SM in high vegetation-covered areas by spectral feature space methods. However, the accuracy of MSMMI was slightly higher than that of the semi-empirical SAR model in estimating SM in bare soil areas. SAR data are more suitable for monitoring SM in vegetation-covered areas, especially in high vegetation-covered areas. When combined with Sentinel-1 and Sentinel-2 to estimate SM, the R of the estimation SM and the measured SM is 0.8886, and the RMSE is 0.0391cm³/cm³, which is obviously better than using optical or SAR data alone. The accuracy of machine learning regression models in SM estimation is significantly higher than that of semi-empirical SAR model and spectral feature space methods. GRNN and RFR algorithms are better than SVR algorithm in estimation accuracy and efficiency. Combined with the traditional four-band spectral features (Blue, Green, Red and Near-infrared) and dual-polarization SAR features (VV, VH and radar incidence angle), SM can be accurately estimated. The R of the estimation SM and the measured SM based on RFR can reach 0.9355, and the RMSE can reach 0.0284cm³/cm³. When combined with optical and SAR data by DNNR algorithm to estimate SM, the RMSE can reach 0.0045cm³/cm³. Combined with Sentinel-1 and Sentinel-2 data, it has great potential to monitor SM at regional scale with spatial resolution of 10m~20m.