田间尺度土壤水分主被动遥感联合反演算法及贝叶斯优化

Real-time and dynamic monitoring of soil moisture at the field scale is of important significance for obtaining crop growth information, drought forecasting and warning, farmland irrigation and other agricultural scientific research fields. Nowadays, the method of retrieving surface soil moisture by using optical remote sensing cannot accurately describe the scattering characteristics of actual features. Active-passive (radar and optical) remote sensing has shown good development prospects for improving the accuracy of soil moisture retrieval. This project intends to propose an active-passive remote sensing to retrieve field-scale soil moisture and Bayesian optimization of its parameters. The details include 1) according to the surface coverage types and vegetation canopy water content values, the segmentation features are selected and an object-oriented segmentation method is carried out on high-resolution remote sensing images. The spatial relationship between adjacent pixels of the image is considered and the segmentation object is determined as the basic unit of soil water retrieval. 2) combined with optical and radar image data, the retrieval mechanism model of soil moisture under different surface types is constructed and the implicit relationship between soil moisture and backscatter coefficient is studied, which is helpful to further understand the cooperative retrieval mechanism of optical and radar remote sensing and provide technical support for soil moisture retrieval in the coverage of complex surface and multi-vegetation region. 3) by establishing Bayesian optimization algorithm and combining with the soil moisture retrieval mechanism model, the optimal estimations of model parameters of each object are calculated and the estimated values of soil moisture of each pixel in the object are obtained. The results are compared with the observed data of the ground station and the field measured data. Finally, a perfect retrieval method and theoretical system for soil moisture on the field scale are established.

田间尺度土壤水分的实时、动态监测对获取农田作物生长信息、旱情预测预报以及农田灌溉等农业科学研究有着十分重要的指导意义。目前仅利用光学遥感手段反演地表土壤水分方法无法准确描述实际地物的散射特性，而主被动（雷达和光学）遥感对于提升土壤水分反演精度已显示出良好的发展前景。本项目拟提出一种主被动遥感联合反演田间尺度土壤水分算法并对其参数开展贝叶斯优化，包括：1）对高分辨率遥感影像进行面向对象分割，考虑影像相邻像元的空间关系，并将分割对象确定为土壤水分反演的基本单元；2）构建不同地表覆盖类型土壤水分反演机理模型，研究土壤水分与后向散射系数之间隐式关系并生成模拟数据集，深入理解光学和雷达遥感联合反演机制；3）利用贝叶斯优化算法计算每个对象的模型参数最优估计，获取对象内各像素土壤水分估测值，并将研究结果与地面站点观测资料及野外实测数据进行比较验证，建立完善的应用于田间尺度地表土壤水分反演方法与理论体系。

土壤水分参与陆地与大气能量交换，是陆地表面蒸散、水分运移以及碳循环过程中的重要因子，对水文过程以及气候变化有着重要的调节作用，对土壤水分进行动态、实时监测对于我国开展农田灌溉、旱情预测预报以及作物产量估计等农业研究具有十分重要的指导意义。本项目将光学影像良好的空间拓扑关系应用到高分辨率雷达数据中，建立适用于土壤水分反演的高分辨率光学和雷达遥感数据面向对象分割特征数据，利用光学和雷达遥感反演地表土壤水分机理模型建立土壤水分与干旱指数或后向散射系数之间的关系，结合优化算法获取模型参数最优估计，提高不同地表类型覆盖下地表土壤水分反演精度，实现不同地表类型覆盖下土壤水分反演和典型区域的应用示范。本项目研制出高分辨率光学和雷达遥感影像分割对象特征数据，建立条件植被干旱指数和归一化温度干旱指数等光学遥感指数反演植被覆盖区土壤水分，构建低矮植被后向散射模型以及改进的作物散射模型研究不同地表覆盖类型下土壤水分的估算，最终构建改进粒子群神经网络优化算法进一步优化模型参数，实现不同地表覆盖类型下土壤水分的高精度反演研究。.依托本项目共发表SCI论文3篇，其中两篇二区top论文，分别被Remote Sensing和Photogrammetric Engineering & Remote Sensing等遥感主流期刊收录，另外获批国家发明专利一项，其研制内容已得到中国气象局国家气象信息中心的认可，并已在农业灌溉、干旱灾害监测等业务化系统中部署应用。同时联合培养博士研究生一名，硕士研究生2名，项目负责人受邀担任Frontiers in Artificial Intelligence：AI in Food, Agriculture and Water期刊客座编辑；受邀组织Remote Sensing期刊下专刊“Remote Sensing for Soil Moisture and Vegetation Parameters Retrieval”，并在国际学术会议上做口头学术报告。