基于作物生长模型与机器学习算法的区域冬小麦估产

Regional Winter⁃wheat Yield Estimation Based on Coupling of Machine Learning Algorithm and Crop Growth Model

为精准、高效、实时地实现区域冬小麦产量估算,以河南省鹤壁市淇县桥盟乡石桥村为研究区,基于分辨率10 m的Sentinel-2多时相光学遥感影像,利用集合卡尔曼滤波(Ensemble Kalman filter,EnKF)算法同化PROSAIL辐射传输模型反演的多期叶面积指数(Leaf area index,LAI)到PyWOFOST作物生长模型中实现一定数量不同长势单点产量的估测,最后利用建立的机器学习模型和面域数据反演区域冬小麦产量,实现作物生长模型与机器学习算法的应用耦合及一种新的区域冬小麦估产模式。研究基于Sobol参数敏感性分析法量化对贮藏器官总干重质量(Total dry weight of storage organs,TWSO)与LAI_(max)的敏感性参数,并基于反演的多期LAI和粒子群优化(Particle swarm optimization,PSO)算法优化与LAI_(max)相关的TDWI、TBASE、CVS、CVL敏感性参数,将其输入到PyWOFOST模型中,利用EnKF算法和时序LAI数据调整对TWSO相关的AMAXTB1、TDWI、TSUMEM、CVO敏感性参数,实现单点产量的估算;与实测单点产量相比,该方法估算的R2、RMSE、MAE、Bias分别为0.8665、468.64 kg/hm^(2)、385.70 kg/hm^(2)和103.08,为建立随机森林回归(Random forest regression,RFR)区域估产算法提供准确的单点产量训练数据。针对研究区(309.32 hm^(2)),基于不同长势人工样点产量数据建立的RFR区域估产算法,区域估产精度为99.44%,每公顷算法运行用时1.55 s;应用EnKF算法同化多时期面域LAI到PyWOFOST作物生长模型中的区域估产精度为89.01%,每公顷算法运行用时约0.47 h;耦合PyWOFOST作物生长模型与RFR机器学习算法的区域估产精度达到95.58%,每公顷算法运行用时8.85 s(训练数据的单点产量计算占总时长约81.35%),显著降低机器学习算法所需的人工成本和同化变量过程计算的时间及算力成本。研究结果为准确、快速的大区域作物估产提供理论支持和技术参考。

To realize the regional winter wheat yield estimation accurately,efficiently and in real⁃time,Shiqiao Village,Qi County,Hebi City,Henan Province,was taken as the study area.The ensemble Kalman filter(EnKF)was used to assimilate the time⁃series leaf area index(LAI),which were estimated by the PROSAIL radiation transfer model,into PyWOFOST crop growth model to estimate a certain number of winter wheat site yield points with different growth.And those site yield points provided training data for random forest regression(RFR)algorithm to establish machine learning model.Finally,the established machine learning model and the time⁃series optical remote sensing images of Sentinel 2 with 10 m resolution were used to estimate the regional winter wheat yield,so as to realize the application of coupling crop growth model and machine learning algorithm,and establish a new regional winter wheat yield estimation mode.Based on Sobol parameter sensitivity analysis algorithm,the sensitivity parameters of TWSO and LAI_(max) were quantified.The TDWI,TBASE,CVS and CVL sensitivity parameters related to LAI_(max) were optimized by time⁃series LAI data and particle swarm optimization(PSO)algorithm.And inputting them into the PyWOFOST model,using the EnKF algorithm and time⁃series LAI data to adjust the AMAXTB1,TDWI,TSUMEM,and CVO sensitivity parameters of TWSO to improve the accuracy of the single⁃point yield estimation.Compared with the site yield points,the R 2,RMSE,MAE,and Bias of estimation were 0.8665,468.64 kg/hm ^(2),385.70 kg/hm ^(2) and 103.08,respectively,providing accurate site points yield of training data for establishing the RFR region yield estimation algorithm.