遥感与作物生长模型数据同化应用综述

遥感是获取大面积地表信息最有效的手段，在农业资源监测、作物产量预测中发挥着不可替代的重要作用；作物生长模型能够实现单点尺度上作物生长发育的动态模拟，可对作物长势以及产量变化提供内在机理解释。遥感信息和作物生长模型的数据同化有效结合二者优势，在大尺度农业监测与预报上具有巨大的应用潜力。该文系统综述了遥感与作物生长模型的同化研究，概述了遥感与作物生长模型数据同化系统的构建，在归纳国内外研究进展的基础上，总结了当前主流同化方法的特点以及在不同条件下的同化效果。进而具体分析影响同化精度的关键环节，明确了相关科学概念，并相应指出改善精度的策略或者方向。最后从多参数协同、多数据融合、动态预测、多模型耦合以及并行计算环境5个方面展望了遥感与作物生长模型数据同化的未来研究重点和发展趋势，同时结合农业应用现实需求，介绍一种数据同化与集合数值预报结合的应用框架，为大区域、高精度同化研究提供新的思路与借鉴。

Review on data assimilation of remote sensing and crop growth models

Abstract: Data assimilation technology, which can combine the advantages of remote sensing and crop growth models, has great potential in large-scale application of agricultural monitoring and yield forecasting. This review included 5 parts. And the first part introduced the framework of data assimilation system of crop growth model and remote sensing. The data assimilation system contained 3 basic components: dynamic model, observation data and assimilation algorithm. Taking the WOFOST model as an example, a schematic representation of assimilating remotely sensed data into a crop model was shown. The second part summarized the progress of data assimilation of crop growth model and remote sensing. The parameter optimization methods based on cost function and the sequential filtering methods based on estimation theory were two major groups of modern data assimilation strategies. The main difference between the two groups was that each subsequent observation for sequential filtering assimilation would only influence the change nature of the model from the current state; in contrast, parameter optimization methods adjusting the estimation using all of the available observations throughout the assimilation window. In general, MODIS data was the most commonly used remotely sensed data for large regional assimilation research, and data of Landsat TM, ETM+ and OLI were the major remotely sensed data used at regional scale. General models, like WOFOST, CERES, etc. were most widely used in agricultural data assimilation researches. The main object of these researches was food crops such as wheat, corn and rice. LAI (leaf area index) was most commonly used as the assimilation variable linking remote sensing and crop models. In addition, a number of studies found that time series of reflectance, vegetation index or backscattering coefficient could be directly assimilated into a coupled crop growth-radiative-transfer model to avoid the process of regional LAI retrieval. In general, yield estimation and forecast was the most important application. The third part discussed some key aspects affecting the assimilation accuracy, including 5 parts: 1) The pixel size for assimilation, which depended mainly on the specific application. However, heterogeneous, smallholder farming environments presented significant challenges for remotely sensed data assimilation for crop yield forecasting, as field size within these highly fragmented landscapes was often smaller than the pixel size of remote sensing products that were freely available. 2) The uncertainty of remote sensed parameter inversion, which needed to be quantitatively evaluated to ensure the accuracy of data assimilation. 3) The uncertainty of crop growth models, which caused by model structure, model parameters and weather driven data. 4) Data assimilation strategies and linking parameters. Two main data assimilation strategies, parameter optimization and sequential filtering methods, both had pros and cons. Therefore, more effective assimilation algorithms still needed to be developed. 5) The scale effect. Due to the variability of land cover and the complexity of the crop planting pattern in agricultural landscapes, the scale mismatch between the remotely sensed observations and the state variables of crop growth models remained a difficult challenge. The fourth part summarized the research trend of data assimilation for crop growth model and remote sensing. It included 5 directions: 1) from single assimilated parameter to multiple ones. 2) from single remotely sensed data to multiple ones, especially the combination of optical remote sensing and SAR(synthetic aperture radar). 3) from monitoring to forecasting. Based on former researches, an application framework combining data assimilation and numerical prediction was proposed. 4) from single crop growth model to multi-crop model coupling. 5) from single machine system to the high-performance parallel computing system, especially considering the recent advances in Google Earth Engine.