Cui Ying, Liu Suhong, Li Xingang, Geng Hao, Xie Yun, He Yuhua
Faculty of Geographical Science, Beijing Normal University, Beijing, China.
Beijing Key Laboratory of Environmental Remote Sensing and Digital Cities, Beijing Normal University, Beijing, China.
Front Plant Sci. 2022 Jun 23;13:915109. doi: 10.3389/fpls.2022.915109. eCollection 2022.
Accurate yield estimation at the regional scale has always been a persistent challenge in the agricultural sector. With the vigorous emergence of remote sensing land surface observations in recent decades, data assimilation methodology has become an effective means to promote the accuracy and efficiency of yield estimation by integrating regional data and point-scale crop models. This paper focuses on the black soil area of Northeast China, a national strategic grain production base, applying the AquaCrop crop growth model to simulate the fractional vegetation cover (FVC) and maize yield from 2000 to 2020 and then forming a reliable FVC optimization dataset based on an ensemble Kalman filter (EnKF) assimilation algorithm with remote sensing products. Using the random forest model, the regression relationship between FVC and yield was established from the long-term time series data, which is crucial to achieve better yield estimation through the optimized FVC. The major findings include the following: (1) The R of the assimilated FVC and maize yield can reach 0.557. (2) When compared with the local statistical yield, our method reduced the mean absolute error (MAE) from 1.164 ton/ha (based on GLASS FVC products) to 1.004 ton/ha (based on the calibrated AquaCrop model) and then to 0.888 ton/ha (the result after assimilation). The above results show that we have proposed a yield estimation method to provide accurate yield estimations by combining data assimilation and machine learning. This study provided deep insights into understanding the variations in FVC and revealed the spatially explicit yield prediction ability from the time series land surface parameters, which has significant potential for optimizing water and soil resource management.
区域尺度上的精确产量估计一直是农业领域持续面临的挑战。近几十年来，随着遥感地表观测的蓬勃兴起，数据同化方法已成为通过整合区域数据和点尺度作物模型来提高产量估计准确性和效率的有效手段。本文聚焦于中国东北的黑土区，这是国家战略性粮食生产基地，应用AquaCrop作物生长模型模拟2000年至2020年的植被覆盖度（FVC）和玉米产量，然后基于集合卡尔曼滤波（EnKF）同化算法与遥感产品形成可靠的FVC优化数据集。利用随机森林模型，从长期时间序列数据中建立了FVC与产量之间的回归关系，这对于通过优化后的FVC实现更好的产量估计至关重要。主要研究结果如下：（1）同化后的FVC与玉米产量的R值可达0.557。（2）与当地统计产量相比，我们的方法将平均绝对误差（MAE）从1.164吨/公顷（基于GLASS FVC产品）降至1.004吨/公顷（基于校准后的AquaCrop模型），然后降至0.888吨/公顷（同化后的结果）。上述结果表明，我们提出了一种通过结合数据同化和机器学习来提供精确产量估计的方法。本研究为理解FVC的变化提供了深刻见解，并揭示了从时间序列地表参数进行空间明确的产量预测能力，这在优化水土资源管理方面具有巨大潜力。
