基于多源遥感数据同化的冬小麦产量估测研究

The fusion of multi-source remote sensing (RS) data with different spatial and temporal resolutions can make up the deficiency of using single RS data source, which is used for estimating crop yield estimation based on data assimilation method. The leaf area index (LAI) and soil moisture (SM) are selected as the assimilation variables in this project, and the Henan province is chosen as the study area. Then the RS data fusion method is used to fuse MODIS, Landsat and sentinel-2 data, and a time series of 30-m-scale LAI is retrieved based on the fused RS data. The Landsat-derived vegetation temperature condition index (VTCI) is applied for obtaining surface SM, which is combined with the SM retrieved from Sentinel-1 SAR data for establishing 30-m-scale SM with high temporal resolution. Then the remotely sensed LAI and SM are assimilated into the CERES-Wheat model for estimating the yield of winter wheat in Henan province using the four-dimensional variational (4DVAR), ensemble Kalman filter (EnKF) and particle filter (PF) approaches, respectively. Finally, the field-measured wheat yield is used for validating and evaluating the accuracies of wheat yield estimation approach by data assimilation. Therefore, based on the previous researches of estimating wheat yield by assimilating single RS data source, the characteristics and innovation of this project are to retrieve and assimilate LAI and SM precisely on ths basis of multi-source remote sensing data. And the study achievements of this project will be very important and useful for regional crop yield estimation.

融合不同时空分辨率的多源遥感数据能够弥补单一遥感数据源在遥感数据同化估产研究中的不足。本项目以河南省为研究区域，以叶面积指数（LAI）和土壤水分（SM）为同化变量，基于遥感数据融合方法融合MODIS、Landsat和Sentinel-2数据进而反演高时间分辨率的30 m LAI；通过Landsat数据反演条件植被温度指数（VTCI）进而反演SM，结合Sentinel-1 SAR数据反演的SM以构建高时间分辨率的30 m SM；采用四维变分、集合卡尔曼滤波和粒子滤波算法同化遥感反演的和CERES-Wheat模型模拟的LAI、SM以估测区域冬小麦单产；利用田间实测小麦单产对相关的遥感数据同化估产方法进行精度验证和评价。本项目的特色和创新是基于前期单一遥感数据源同化估产的研究成果，开展基于多源遥感数据的LAI、SM的高精度反演和同化估产研究，以期研发服务于区域农业生产的作物估产方法。

作物生长模型和遥感数据各有优缺点，将作物生长模型引入遥感作物估产，实现两者的优势互补，是促进大面积冬小麦长势监测和产量估测、预测向机理化和精确化方向发展的有效途径。本项目结合多源遥感数据和CERES-Wheat模型进行了河南省冬小麦产量估测。首先，基于时间序列Sentinel-2数据和冬小麦物候特征构建小麦制图指数，结合自动阈值法对河南省冬小麦种植区域进行提取；然后，获取了2014—2019年冬小麦生长季35个农业气象站点的逐日气象数据、土壤参数、田间管理参数和冬小麦生长数据，对CERES-Wheat模型进行了标定和验证；最后，采用CERES-Wheat模型模拟的小麦叶面积指数（leaf area index，LAI）和产量以及田间实测的小麦LAI和产量，分别训练了长短期记忆网络（long short-term memory，LSTM）、一维卷积神经网络（one-dimensional convolutional neural network，1-D CNN）和随机森林（random forest，RF）模型，通过多源遥感数据反演LAI，基于训练的模型和遥感反演LAI进行了河南省冬小麦产量估测，并进行精度对比。结果表明，结合指数法和自动阈值法能够降低冬小麦制图对样本数据的依赖，实现小麦种植区域的自动提取；标定的CERES-Wheat模型能够准确地模拟河南省冬小麦的生长发育过程，模拟的小麦LAI和产量精度较高；基于LSTM模型的小麦估产精度高于基于1-D CNN模型和RF模型的小麦估产精度，说明LSTM模型能够有效地从时序遥感数据中获取小麦的物候信息从而用于产量估测；基于CERES-Wheat模型模拟的LAI和产量训练LSTM模型，然后采用田间实测的LAI和产量进一步优化LSTM模型，所获得的LSTM模型的估产精度高于仅采用CERES-Wheat模型模拟数据或田间实测数据训练的LSTM模型的估产精度。CERES-Wheat模型模拟数据能够准确地表征田间尺度小麦的生长发育特征，而遥感数据能够揭示区域尺度小麦的生长状况，因此，结合两种数据源在小麦估产方面更有优势。本项目的研究成果能够为大面积遥感作物估产研究提供重要的理论基础。