Knowledge Resource Center for Ecological Environment in Arid Area
| 土壤水分微波遥感模型与反演的不确定性研究 | |
| 其他题名 | Quantifying Uncertainties in Microwave Remote Sensing Modeling and Inversion of Soil Moisture |
| 马春锋 | |
| 出版年 | 2016 |
| 学位类型 | 博士 |
| 导师 | 李新 |
| 学位授予单位 | 中国科学院大学 |
| 中文摘要 | 土壤水分(Soil Moisture, SM)是连接水循环、能量循环和生物地球化学循环的关键变量，在陆地水循环中扮演着极其重要的角色。因此，对土壤水分的准确估计是众多学科领域（如水文学、气象学、农学等）研究的热点问题。微波遥感以其独特的穿透性及全天时、全天候的观测优点，在土壤水分估计中发挥着举足轻重的作用。自上世纪70年代以来，基于微波遥感的土壤水分研究取得了突飞猛进的发展，一是微波遥感器的不断更新和卫星观测计划的发起实施，形成了多角度、多极化、多尺度的立体对地观测网；二是遥感正向模型和土壤水分反演算法的不断改进，使模拟和反演的精度不断提升；三是土壤水分观测试验及观测网络的实施和建立，提供了丰富的地面验证数据集。然而，目前对土壤水分微波遥感模型和反演的不确定性认识不足，导致如下三方面问题：(1) 采用局部敏感性分析方法，无法全面刻画地表参数及其相互作用所引起的微波遥感观测的不确定性；(2) 对微波散射模型在像元尺度上的不确定性缺乏必要的探索，从而导致模型模拟与遥感观测存在明显的差异；(3) 反演算法以确定性的方法为主，仅能给出土壤水分的点估计，无法获取土壤水分在概率框架下的后验估计，更无法量化反演的不确定性。鉴于此，本文以探索并量化土壤水分微波遥感的不确定性为目标，围绕该目标开展了如下三方面的研究工作：（1） 开展地表参数的全局敏感性分析，量化参数不确定性采用扩展傅里叶振幅敏感性分析算法(Extended Fourier Amplitude Sensitivity Test, EFAST)、Sobol’算法、Delta Test (DT)、Derivative-based Global Sensitivity Measurements (DGSM) 等多种全局敏感性分析算法，分析了微波观测量对地表参数的敏感性。被动微波遥感的地表参数敏感性分析以高级积分方程模型（Advanced Integral Equation Model, AIEM）的发射模型及土壤有效温度参数化为基础，采用EFAST定量评估了发射率及亮度温度对地表参数的敏感性。同时，采用Sobol’、DGSM、DT算法对分析结果进行了强化对比。多种分析算法的结果一致表明土壤水分、粗糙度均方根高度（Root Mean Square of surface Height , RMSH)和相关长度（Correlation lengt, CL:)最为敏感，三者对发射率和亮温的主敏感性指数之和分别为0.92和0.95。同时发现：①参数的概率分布对其敏感性和模型输出的影响甚微；②随入射角和频率的变化，参数敏感性发生明显的变化，在低频率、大入射角条件下，土壤水分最为敏感；③地表自相关函数明显影响参数的敏感性，高斯函数突出了土壤水分的敏感性，而指数函数则突出了粗糙度的敏感性；④ V极化发射率及亮温对SM的敏感性略大于H极化，而H极化则对RMSH较敏感，另外，微波极化指数对RMSH的敏感性明显大于对其余任何参数。主动微波遥感的参数敏感性分析以AIEM后向散射模型为基础，采用EFAST量化了后向散射系数及其组合（双极化后向散射系数差值VV-HH或比值VV/HH、双角度后向散射系数差θ1-θ2、双频后向散射系数差f1-f2等）对地表参数及其相互作用的敏感性。结果表明， RMSH的敏感性远大于其它参数，其次是SM和CL；参数的取值范围、相关函数、入射角、频率、极化等对参数敏感性具有明显的影响，低频条件下，参数的主敏感性最强，随频率增大，其相互作用增强；L波段VV-HH对RMSH的敏感性明显增强，而VV/HH对SM和CL的敏感性增强，θ1-θ2在C波段VV极化下对RMSH的敏感性增强，而其在L波段VV极化下则对SM的敏感性几乎可以消除； f1-f2导致参数之间的相互作用增强。总之，地表参数的全局敏感性分析克服了局部敏感性分析的不足，量化了参数敏感性及其相互作用。定量解释了土壤水分估计对微波传感器的配置参数（波段、极化、入射角等）的需求，为探索土壤水分反演的潜在数据源提供了基础。定量表达了参数敏感性所导致的微波散射/发射模型的不确定性，为土壤水分反演的不确定性量化奠定了基础。（2） 开展微波散射模型比较与评估，衡量模型不确定性微波散射模型是理解地表散射过程和开展土壤水分反演的重要工具，系统评估散射模型对土壤水分主动微波遥感至关重要。本论文对9个微波后向散射模型进行了比较与评估。评估方案涉及理想试验及实测试验，前者分别采用等步长采样和蒙特卡罗采样，分析后向散射系数对土壤水分的变化响应和对土壤水分、粗糙度的综合响应。后者采用SMAPVEX12试验中 L波段机载雷达观测与模型输出对比，考察模型在宏观尺度上的不确定性。1）模型对土壤水分的响应分析表明，除Dubois模型外，大多数散射模型对土壤水分的变化响应较为一致。模型对SM，RMSH和CL的响应分析定量刻画了敏感参数及其相互作用对雷达观测的综合贡献。2）模型输出的概率分布呈现明显的负偏态分布，且其期望输出与实际输出的统计值存在明显差异，揭示了微观上完备的散射模型在宏观尺度上的不确定性。3）随地表参数范围向其均值方向收缩（即参数均值不变，范围缩小），模型输出值的分布范围并未随之收缩，而是其概率峰值向低值区平移，表明了具有相同期望、不同范围的地表参数引起模型输出的实际分布与其期望分布的差异。4）模型模拟与L波段雷达观测的对比试验表明模拟值与观测值之间存在较大的差异，初步分析，该差异与粗糙度的尺度效应有关。这就意味着微观尺度的微波散射模型应用于像元尺度时，需要发展与尺度有关的粗糙度等效参数。（3） 发展概率反演算法，量化反演的不确定性基于贝叶斯-蒙特卡罗马尔可夫链技术，发展了土壤水分的贝叶斯概率反演方法，该方法在贝叶斯理论框架下反演得到了土壤水分的后验概率分布，量化了反演的不确定性（定义土壤水分后验分布的标准差为反演的不确定性，其值越大，则不确定性越大），并通过最大似然估计代表土壤水分的最优估计值，从而提高土壤水分的估计精度。基于该算法，分别采用TerraSAR-X (TSX) 及航空Polarimetric L-band Multi-beam Radiometer (PLMR)观测在黑河中游开展了反演试验。在TSX的土壤水分反演中，假设壤水分、粗糙度和植被含水量（Vegetation Water Content, VWC）服从均匀分布，分别以AIEM和AIEM-WCM（Water Cloud Model, 水云模型）的模拟值和高分辨率TSX双通道后向散射系数（HH，VV）构造似然函数。基于以上先验信息和似然函数，采用Metropolis-Hastings（M-H）算法在贝叶斯理论的框架下反演得到了SM、RMSH、CL和VWC的后验概率密度函数，通过分析土壤水分等参数概率分布的统计特征，揭示了反演的不确定性。表明其不确定性为0.10-0.12 m3/m3。通过最大似然估计获得了土壤水分的最优估计值，并利用黑河流域生态水文过程综合遥感联合观测试验（HiWATER）的实测土壤水分和植被含水量对反演结果进行了验证。表明土壤水分最优估计值能够较好地反映地表的实际状况，SM在裸土和植被覆盖地表的反演误差RMSE分别为0.045和0.047 m3/m3，VWC的RMSE为0.45kg/m2。基于航空PLMR的土壤水分反演中，首先根据前人的参数敏感性分析结果，确定了SM、Hr（有效粗糙度）和VWC为同步反演目标，并确定其初值和取值范围。将贝叶斯概率反演算法进行了改进，算法中采用PLMR亮温与对应的L-MEB模拟值构造似然函数，采用M-H算法在贝叶斯理论框架下获得了反演目标的后验概率密度，采用最大似然估计获得了土壤水分的最优估计值。反演中采用了两参数同步估计（2P）及三参数同步估计（3P）的反演策略。2P中首先通过亮温直接估计Hr，然后同步估计SM和VWC。3P中同步估计SM、Hr和VWC。结果表明3P的不确定性较2P的不确定性小，其值分别为0.08和0.10 m3/m3。两者在植被覆盖地表均能取得高精度的反演结果，其RMSE分别为0.025和0.023 m3/m3，但两参数法在荒漠地表的反演结果不理想，主要原因在于Hr的建模引起的不确定性。总之，本论文围绕土壤水分微波遥感建模和反演的不确定性，定量分析了地表参数的敏感性，评估了微波散射模型的不确定性，开展了土壤水分的概率反演，量化了反演的不确定性。取得的主要结论有：1）地表参数及其相互作用对微波散射/发射的敏感性是导致土壤水分微波遥感不确定性的重要来源，需要从全局敏感性的角度量化；2）微波散射模型在微观尺度上表现出良好的一致性，但在宏观尺度上与雷达观测存在较大差异，需要发展宏观尺度上等效的模型参数缩小该差异；3） 在概率反演框架下开展土壤水分估算能够揭示反演的不确定性，并提高反演精度。 |
| 英文摘要 | Soil moisture (SM), a key state variable which links water cycle, energy cycle and biogeochemistry cycle, plays very important roles in terrestrial water cycle. As a consequence, an accurate estimation of SM is of significance in many disciplines, such as hydrology, weather forecasting, agriculture and global climate change. Microwave remote sensing plays an important role in estimation of SM due to its distinguished ability in penetrating vegetation and observing the earth under variable weather conditions.Since 1970s, SM estimation based on microwave remote sensing has been witnessed a rapid progress in several aspects: 1) the creation and time to time update of the microwave remote sensors, as well as the initiation and implementation of earth observation for SM, can provide multi-angular, multi-polarized and multi-scales observations of the earth surface, 2) the development and continuous refining of the forward models and SM retrieval algorithms improve the accuracy and precision of SM estimation, and 3) the instigation and execution of a large number of SM observation experiments and observation network towards the development and validation of SM products derived from the microwave remote sensing, as a result provide very rich datasets for SM investigation. However, there still exists some deficiencies cuased by the lack of understanding on uncertainties in microwave remote sensing of SM:I. The use of local sensitivity analysis (SA) methods cannot comprehensively depict the uncertainties in microwave remote sensing observations caused by parameter sensitivity and their interactions;II. The lack of exploration on uncertainties in microwave scattering models at pixel scale causes larger dispancies between model simulations and remotely sensed observationsIII. Most of retrieval algorithms are deterministic inversion which can only obtain the point estimation of SM, neither obtain the posterior estimation of SM under probabilistic framework nor quantify uncertainties in the SM inversion.The thesis subjected to explore and quantify the uncertainties in microwave remote sensing modeling and inversion of SM research, in which three aspects research work were addressed as following.(1) Global SAs of surface parameters were performed for the quantification of parameter uncertaintiesThe sensitivity of soil surface parameters on microwave backscattering coefficient and emissivity (or brightness temperature, TB) were analyzed using various global SA algorithms, namely Extended Fourier Amplitude Sensitivity Test (EFAST), Sobol’ method, Delta Test (DT) and Derivative-based Global Sensitivity Measurements (DGSM).For the SA of parameter on passive microwave observations, the EFAST (Extended Fourier Amplitude Sensitivity Test) was firstly applied to quantitatively evaluate the sensitivity of surface parameters on emissivity/TB based on emission model of AIEM and parameterization of soil effective temperature. Then, Sobol’, DT and DGSM were used to verify and validate the results of EFAST. The results from all of the algorithms consistently showed that SM, RMSH (RMS of surface height) and CL (correlation length) were the most sensitive parameters, the sum of whose main sensitivity indices is larger than 0.92. We also found that: ① the probability distributions of parameters had no impacts on parameter sensitivity indices and the outputs of the model; ② as the incidence angle and frequency vary, the parameter sensitivities change significantly, and the SM was most sensitive at lower frequency and larger incidence angle; ③ surface auto-correlation functions (ACFs) significantly influenced parameter sensitivities; ④ SM was more sensitive on V than on H polarized emissivity and TB; RMSH was more sensitive on H than on V polarized emissivity and TB; the sensitivity of RMSH was much larger than any other parameter on microwave polarization index. For the SA of parameter on active microwave observations, the sensitivities of backscatter on parameters and their combinatons (VV-HH, VV/HH, θ1-θ2 and f1-f2) were quantitatively evaluated based on the combination of active formulation of AIEM and EFAST. The results showed the RMSH sensitivity is much larger than others, followed by SM and CL. In addition, significant impacts of parameter ranges, ACFs, incidence angle, frequency and polarization on sensitivity measurements were observed. Dual-polarization, dual-incidence angle and dual-frequency combinations were significantly influenced the parameter sensitivities. The RMSH was more sensitive on VV-HH at L-band, and VV/HH enhanced the sensitivities of SM and CL. The SM sensitivity was significantly decreased, and that of the roughness was increased on difference of two incidence angles backscatter at C-band. The SM sensitivity on two incidence angles difference at L-b and VV polarization was nearly eliminated. However, the parameter interactions were enhanced at two frequencies difference.Generally, the use of global SA of surface parameter on microwave observations presented in this thesis overcomes the deficiency of the previous local SAs, quantifying the parameter sensitivities and their interactions, providing basis for potential data sources for SM inversion, and also analyzing the uncertainty in microwave scattering/emission process caused by the parameter sensitivities. Thus, it provides basis for SM inversion based on microwave remote sensing.(2) The backscattering models were compared and evaluated to examine their uncertainties at larger scalesMicrowave backscattering models are important tools for the understanding of surface scattering process and the forward simulating in SM inversion. The present thesis systematically compared and evaluated the 9 most used models. The evaluation approaches involve synthetic and real data, respectively. The former analyzed the responses of backscatter on SM, RMSH and CL using equal step length sampling and Monte Carlo randomly sanpling approaches. The latter directly compares models simulations with L-band radar observation from SMAPVEX12 experiment. The response of model simulations on SM showed that most of the models presented consistent performance except the Dubois model. The synthetic response of models on SM, RMSH and CL quantitatively represents the synthetic contribution of sensitive parameters on radar observations. The distribution of the model output shows significantly positively skewed model, and the peak value of the distributions moved to the smaller value parts as the parameter ranges shrink toward median values. The comparison of model simulations and L-band radar observations showed larger inconsistence, which may result from the incorrect surface roughness at pixel scale. This means that the use of microwave scattering models on a pixel scale may cause uncertainties in the model simulations, which suggests that the scale-dependent effective roughness parameters are dispensable.(3) Probabilistic inversion (PI) was proposed to quantify the uncertainties in SM inversionA SM probabilistic inversion (PI) approach based on Bayesian Monte Carlo Markov Chain technique was proposed, which quantified the uncertainties in SM inversion (We define the standard deviation of SM posterior distribution as uncertainty, the larger value means larger uncertainty) by obtaining the posterior distribution of SM under the framework of Bayes theorem. And the optimal estimators of SM was obtained by Maximum Likehood Estimation (MLE), hence the estimation precision been improved. Based on the Bayesian PI, SM inversion experiments were conducted by using TerraSAR-X (TSX) and airborne PLMR of middle reaches of Heihe River Basin, respectively. The PI of SM based on TSX scheme applied TSX backscatter and AIEM_WCM coupled model to construct cost function, under the basis of the assumption that SM, roughness parameters and VWC (vegetation water content) are uniformly distributed. The posterior estimations of SM, roughness and VWC in the farmland were obtained. The SM inversion uncertainty was revealed by analyzing the stastical characteristic of SM posterior distributions, which shows 0.10-0.12 m3/m3 uncertainty in SM inversion. The RMSEs of retrieved SM were 0.045 and 0.047m3/m3 at bare and vegetated soils, respectively. The RMSE of VWC was 0.45kgm-2.The estimation of SM using PLMR combined L-band Microwave Emission of the Biosphere simulations and multi-angular, dual-polarized TB of PLMR observation to retrieve SM within the revised Bayesian PI, as well as, to analyze the inversion uncertainties. Two inversion strategies (namely, three parameters were simultaneously estimated, 3P, and two parameters were simultaneously estimated, 2P) were applied. It showed both of the two strategies resulted in high precision estimation of SM in vegetated surface, with the RMSE were 0.025 and 0.023m3m-3, respectively. However, the 2P strategy leaded to larger RMSE of the inversion results, the main reason of which may be caused by the incorrect modeling of effective roughness. Generally, the presented thesis subjects to the the quantification of uncertainties in microwave remote sensing modeling and inversion of SM, quantitatively analyzing parameter sensitivities, evaluating the uncertainties in backscattering models and propsosing and estimation approach for SM basen on PI, quantifying the uncertainties in SM inversion. The main conculsions were drawn: 1) Parameter sensitivities and their interaction are important sources of modeling and inversion, which need be quantified from the global sensitivity aspects. 2) Most of backscattering models are consistent each other, but their still exists larger dispancies between model simulations and radar observations, which suggests that the scale-dependent effective roughness parameters are dispensable. 3) The estimation of SM under the PI framework can reveal the uncertainties in SM inversion. |
| 中文关键词 | 微波遥感 ; 土壤水分 ; 不确定性 ; 概率反演 ; 散射模型 ; 发射模型 |
| 英文关键词 | Microwave remote Sensing Soil moisture Uncertainty Probabilistic inversion Scattering models Emission models |
| 语种 | 中文 |
| 国家 | 中国 |
| 来源学科分类 | 地图学与地理信息系统 |
| 来源机构 | 中国科学院西北生态环境资源研究院 |
| 资源类型 | 学位论文 |
| 条目标识符 | http://119.78.100.177/qdio/handle/2XILL650/287717 |
| 推荐引用方式 GB/T 7714 | 马春锋. 土壤水分微波遥感模型与反演的不确定性研究[D]. 中国科学院大学,2016. |
| 条目包含的文件 | 条目无相关文件。 | |||||
| 个性服务 |
| 推荐该条目 |
| 保存到收藏夹 |
| 导出为Endnote文件 |
| 谷歌学术 |
| 谷歌学术中相似的文章 |
| [马春锋]的文章 |
| 百度学术 |
| 百度学术中相似的文章 |
| [马春锋]的文章 |
| 必应学术 |
| 必应学术中相似的文章 |
| [马春锋]的文章 |
| 相关权益政策 |
| 暂无数据 |
| 收藏/分享 |
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。
