多云多雨地区时间序列信息提取应用驱动下的环境星数据预处理技术研究

The HJ multi spectral images offer great potential for multitemporal applications in agriculture because of the satellites' wide swath and short revisit time. And the multitemporal applications of HJ multi spectral images require radiometric corrected, geometric registered and cloud free time series data. However, due to the errors produced in the data acquisition and preprocessing, the multitemporal HJ multi spectral images are often with unsatisfactory radiometric and geometric registration errors, some even are covered with cloud, especially in the cloud-prone and rainy area. On this point, the project focuses on the fine preprocessing and the standard and automatic process pipeline of the multitemporal HJ multi spectral images. Particularly, the geometric and radiometric registration between the time series images and the reference ones and cloud shadow removal are explored. In detail, the thin plate spline, considered as a non-rigid transformation model, is introduced to the geometric registration; content free radiation normalization is used in the radiometric correction; the spatial structure of the pixel values and the local variant spectra of the multitemporal images are both taken into account in the cloud shadow removal algorithm. In practice, the above algorithms are implemented to produce time series images in the application of rice growth monitoring centering on distinguishing early, intermediate and late rice using the leaf area index combined with the prior knowledge of rice phenological characteristics in the Hangzhou-Jiaxing-Huzhou area, Zhejiang Province.

环境星(HJ)多光谱数据能够大面积同步观测、重访周期短，具有较大的多时相应用潜力，而获得辐射、几何一致的时间序列无云影像是其成功应用的基础。受影像本身、处理精度与天气的影响，多时相HJ数据之间存在着较大的辐射、配准误差和云影遮挡现象。鉴于此，本项目着力于解决多云多雨地区时序信息提取应用驱动下的HJ 星数据预处理关键技术，建立HJ 星数据的标准化、自动化处理链条，重点探索多时相数据及其参考数据之间几何、辐射一致性处理和云影去除等方法。运用薄板样条等具有局部改正的非刚性解算模型、内容无关辐射归一化方法，实现影像之间的几何、辐射一致性处理；协同利用多时相影像的灰度值空间结构与光谱局部变异信息，发展云影区域的无缝填充方法；最后，实现相关算法、研发软件，并以浙江省杭嘉湖地区作为实验区，结合水稻物候特征等先验知识，重点根据叶面积指数变化区分早中晚稻，开展水稻种植结构提取的多时相应用示范。

本项目严格按照项目计划开展各项研究。在数据收集与软件平台搭建等工作的基础上，发展了仿射不变量支持的准稠密控制点匹配算法，再运用薄板样条等具有局部改正的非刚性解算模型，实现HJ星影像非均匀形变的校正；通过稳定时间序列建模实现时相不变特征提取，结合整体最优化的归一化策略，实现了长时间序列影像的辐射归一化；在基于影像的云影距离自适应估计实现云影检测基础上，协同利用多时相影像的灰度值空间结构与光谱局部变异信息，发展云影区域的无缝填充方法，实现了无云时间序列影像重建；最后，在实现相关算法的基础上、研发多时相HJ星数据预处理软件，并进行了水稻、小麦农作物种植结构提取实验。针对国产卫星的发展情况，项目进一步拓展了本项目算法在GF系列卫星的数据处理上的应用，同时还利用时间序列环境星星数据开展水体监测等应用。在上述研究基础上，发表了论文和申请专利，同时云影检测和辐射归一化两个关键的国产卫星处理算法已经在数据生产中得到应用，取得了良好的效果。