【作者】 王浩；

【导师】 孙爱华； 魏征；

【作者基本信息】 三峡大学 ， 水利工程（专业学位）， 2022， 硕士

【摘要】 华北平原是我国重要的农牧业生产和商品粮基地,农业用水占总用水量的60%以上。土壤水分可以直接反映土壤的干湿状态,通过利用微波遥感技术,可以准确、实时地获取区域土壤水分,能够有效提升农业干旱监测水平,对于华北平原实施精准灌溉和农业现代化建设有重要的意义。本文以北京市大兴区为典型研究区,利用Sentinel-1、Sentinel-2和Landsat 8影像数据,通过监督分类把研究区分为裸土区和植被覆盖区(春玉米、夏玉米、低矮作物和林地)。在裸土区和植被覆盖区基于水云模型和RBF神经网络模型实现对表层土壤水分的反演,并与以VV极化和VH极化为基础的线性模型反演的表层土壤水分精度进行对比分析。主要结论如下:(1)以Landsat 8影像为基础对大兴区土地利用类型进行监督分类。分类的结果显示:最大似然、神经网络和支持向量机的总体分类精度大于90.53%,Kappa系数大于0.87。平行六面体、最小距离和马氏距离的总体分类精度范围是65.38%~75.12%,Kappa系数范围是0.57~0.68。同时对比分类效果图与实地调查结果,选择最大似然对研究区进行监督分类。(2)去除植被层覆盖影响效果的评价。在用光学影像去除植被覆盖影响的研究中,对于VV极化或VH极化,将指数NDVI计算的植被含水量带入水云模型去除植被层影响的效果优于指数NDWI;将Landsat 8的指数NDVI计算的植被含水量带入水云模型去除植被层影响的效果优于Sentinel-2的指数NDVI;VV极化和VH极化穿透植被层后,VH极化衰减的更多。(3)建立植被覆盖区土壤水分反演模型。用水云模型去除植被覆盖的影响,通过去除植被影响的VV极化和VH极化后向散射系数、VV极化与VH极化的极化差、归一化植被指数NDVI和雷达入射角结合RBF神经网络,构建植被覆盖区土壤水分反演模型,通过验证分析:模型的模拟值和实测值相关性较好r=0.855,均方根误差RMSE=0.024cm~3/cm~3。相对于VV极化线性回归模型相关性(r)提高0.103,均方根误差(RMSE)减小0.034 cm~3/cm~3。相对于VH极化线性回归模型相关性(r)提高0.13,均方根误差(RMSE)减小0.01 cm~3/cm~3。(4)建立裸土区土壤水分反演模型。通过VV极化后向散射系数、VH极化后向散射系数、VV极化与VH极化的极化差和雷达入射角结合RBF神经网络,构建裸土区土壤水分反演模型。通过验证分析:模型的模拟值和实测值相关性较好r=0.796,均方根误差RMSE=0.029cm~3/cm~3。相对于VV极化线性回归模型相关性(r)提高0.044,均方根误差(RMSE)减小0.029cm~3/cm~3。相对于VH极化线性回归模型相关性(r)提高0.071,均方根误差(RMSE)减小0.005cm~3/cm~3。更多还原

【Abstract】 North China Plain is an important animal husbandry production and commodity grain base in China.Agricultural water consumption accounts for more than 60 % of total water consumption.Soil moisture can directly reflect the dry and wet state of soil.By using microwave remote sensing technology,regional soil moisture can be accurately and real-time obtained,which can effectively improve the level of agricultural drought monitoring.It is of great significance for the implementation of precision irrigation and agricultural modernization in North China Plain.Taking Daxing District of Beijing as a typical study area,this paper uses Sentinel-1,Sentinel-2 and Landsat 8 image data to divide the study into bare soil area and vegetation cover area(spring maize,summer maize,low-dwarf crops and woodland)by supervised classification.Inversion of surface soil moisture based on water cloud model and RBF neural network model in bare soil area and vegetation cover area,and compared with the soil moisture accuracy of the linear model based on VV polarization and VH polarization.The main conclusions are as follows:(1)Based on Landsat 8 images,the land use types in Daxing District is supervised Classification results show that: the overall classification accuracy of maximum likelihood,neural network and support vector machine is greater than 90.53%,and the Kappa coefficient is greater than 0.87.The overall classification accuracy of parallelepiped,minimum distance and Mahalanobis distance is between 65.38%-75.12%,and Kappa coefficient is between0.57-0.68.At the same time,the effect map and the field survey results are compared,and the maximum likelihood is finally selected to supervised classification of the study area.(2)Evaluation of the effect of removing vegetation cover.Study on removal of vegetation cover by optical image,For VV polarization or VH polarization,the effect of bringing the vegetation water content calculated by NDVI into the water cloud model to remove the influence of vegetation layer is better than that of NDWI;Bringing the vegetation water content calculated by Landsat 8’s index NDVI into the water cloud model to remove the effect of vegetation layer is better than Sentinel-2’s index NDVI;After VV and VH polarization penetrate the vegetation layer,VH polarization decays more.(3)Establishment of soil moisture inversion model in vegetation coverage area.The influence of vegetation cover is removed by the water cloud model,by removing the VV polarization and VH polarization backscatter coefficient,the polarization difference between VV polarization and VH polarization,the normalized vegetation index NDVI,the radar incident angle combined with RBF neural network Constructing a soil moisture retrieval model in vegetation covered area.Through verification analysis,the correlation between the simulated value and the measured value of the model was good(r =0.855,RMSE = 0.024 cm~3/cm~3).Compared with the VV polarization linear regression model,the correlation(r)increased by 0.103,and the root mean square error(RMSE)decreased by 0.034 cm~3/cm~3.Compared with the VH polarization linear regression model,the correlation(r)increased by0.13,and the root mean square error(RMSE)decreased by 0.01 cm~3/cm~3.(4)Establishment of soil moisture inversion model in bare soil area.The soil moisture inversion model in bare soil area was constructed by VV polarization backscattering coefficient,VH polarization backscattering coefficient,polarization difference between VV polarization and VH polarization,radar incident angle combined with RBF neural network.Through verification analysis,the correlation between the simulated value and the measured value of the model was good(r =0.796,RMSE = 0.029 cm~3/cm~3).Compared with the VV polarization linear regression model,the correlation(r)increased by 0.044,and the root mean square error(RMSE)decreased by 0.029 cm~3/cm~3.Compared with the VH polarization linear regression model,the correlation(r)increased by 0.071,and the root mean square error(RMSE)decreased by 0.005 cm~3/cm~3.更多还原