EnKF算法在WOFSOT模型与MODIS数据同化中的应用——以许昌烟叶为例

Application of the Ensemble Kalman Filter (EnKF) Algorithm in the Assimilation of WOFOST Model and MODIS Data—A Case Study of Tobacco in Xuchang

摘要: 遥感与作物模型结合可以实现区域范围上对作物生长发育的模拟，进而应用于区域产量预测和生长监测，有利于指导和决策烟叶相关生产活动。本研究探讨了WOFOST模型与遥感结合在河南优质产烟区许昌地区的应用。基于2021、2022和2023三年气象数据、烟叶产量统计数据、实测土壤数据，以及生长发育期可用MODIS数据，在WOFOST烟叶模型本地化的基础上，利用2023年MODIS Terra和Aqua的LAI数据集对许昌烟叶种植区域进行提取，将遥感数据采用EnKF算法以LAI为中间参数同化WOFOST模型，对烟叶生长发育进行模拟，实现了遥感数据与烟叶模型结合的许昌市烟叶区域模拟。结果表明：(1) 同化后的模型能够适用于许昌市烟叶的生长模拟与产量预测，决定系数R² = 0.979，拟合度指数d = 0.96，均大于0.9，NRSME = 0.139，误差范围在10%~20%，属于较小误差，为可接受范围，模拟效果较好。(2) 同化次数直接影响同化效果，EnKF同化算法的最佳同化次数为50次左右；(3) 利用集合卡尔曼滤波算法，以叶面积指数为结合点，实现了基于遥感数据与作物模型同化的遥感产量估测，校准之后的模型精度平均误差从369 kg·hm⁻²降低到62 kg·hm⁻²，估产精度从67.3%提高到94.5%，表明遥感与作物模型同化是一种有效的提高作物估产和产量预测精度的方法。

Abstract: The integration of remote sensing and crop models enables the simulation of crop growth and development at a regional scale, which can be applied to regional yield prediction and growth monitoring, thereby supporting decision-making and management in tobacco production. This study explores the application of integrating the WOFOST model with remote sensing in Xuchang, a high-quality tobacco-producing region in Henan Province. Based on meteorological data, tobacco yield statistics, measured soil data from 2021 to 2023, and available MODIS data during the crop growth period, the WOFOST tobacco model was first localized. Using the 2023 MODIS Terra and Aqua LAI datasets, tobacco planting areas in Xuchang were extracted. Remote sensing data were assimilated into the WOFOST model using the Ensemble Kalman Filter (EnKF) algorithm, with leaf area index (LAI) as the intermediate variable, to simulate tobacco growth and development. This approach enabled the regional simulation of tobacco growth in Xuchang by integrating remote sensing data with the crop model. The results show that: (1) the assimilated model is suitable for simulating tobacco growth and predicting yield in Xuchang, with a coefficient of determination (R²) of 0.988 and an index of agreement (d) of 0.96, both exceeding 0.9, and an NRMSE of 0.139. The error range is between 10% and 20%, indicating relatively small and acceptable errors and good simulation performance. (2) The number of assimilation iterations directly affects the assimilation performance, with the optimal number of EnKF assimilation iterations being around 50. (3) By using the Ensemble Kalman Filter algorithm and taking LAI as the linking variable, remote sensing-based yield estimation through the assimilation of remote sensing data and the crop model was achieved. After calibration, the average model error decreased from 369 kg·hm⁻² to 62 kg·hm⁻², and the yield estimation accuracy improved from 67.3% to 94.5%, demonstrating that the assimilation of remote sensing data with crop models is an effective method for improving crop yield estimation and prediction accuracy.