Regional Winter-wheat Yield Estimation Based On The Coupling Of Machine Learning Algorithm And Crop Growth Model

Accurate and efficient forecasting of regional winter-wheat yield estimation has been significance for ensuring China’s food security,sustainable development of agricultural,optimization of planting structure and digital agriculture construction.When assimilating remote sensing information with high temporal and spatial resolution and crop growth model method was applied to estimate regional crop yield,higher level of computer performance and excessive time cost are needed due to the limitation of computing power.In recent years,the machine learning methods have been developed rapidly and have strong nonlinear fitting ability and high computational efficiency.It is very necessary to couple machine learning algorithm with crop growth model to improve the accuracy and computational efficiency of regional winter-wheat yield forecasting based on remote sensing information.Taking Hebi City Henan Province as the research area,the main work and conclusions of this dissertation were as follows:(1)With the aim to rapidly and efficiently identify winter wheat by using remote sensing data,an object-oriented method based on Google Earth Engine(GEE)platform and random forest(RF)algorithm were used by integrating the time-series Sentinel-1SAR backward scattering coefficient and the Sentinel-2 spectral features,vegetation index features and texture features of 10m spatial resolution.The result showed that the highest overall accuracy of classification and Kappa coefficient reached 96.33%,and 0.95 at the Simple Non-iterative Clustering(SNIC)segmentation scale of 9,and the Gray-level Co-occurrence Matrix(GLCM)neighborhood value of 4,as well as the first principal components of the 7 texture features.Compared with the statistical area of winter wheat planting area in Hebi city from 2019 to 2021 in the statistical Yearbook of Henan Province,the accuracy of winter wheat planting area identified by remote sensing reached more than 95%and the winter wheat distribution was also suitable with the landform characteristics of Hebi City.(2)Leaf Area Index(LAI)and Soil Moisture(SM)were selected as assimilation variables.Based on Sentinel-2 optical data,using PROSAIL radiation transfer model and MLR machine learning model estimate LAI.The RFR machine learning algorithm,Sentinel active and passive remote sensing data were used to estimate SM.The results showed that the accuracy of the PROSAIL estimating LAI was higher than the MLR machine learning model and the R~2,RMSE,MAE,Bias reached 0.86,0.72,0.57,0.09.The result of estimating LAI accuracy was suitable for the assimilation requirement.RFR algorithm and Sentinel active and passive remote sensing data was used to estimate SM and the R~2,RMSE,MAE,Bias reached 0.51,0.08,0.07,0.01.The results of SM inversion were complied with rainfall events and local actual irrigation conditions.(3)Aiming at the calibration of sensitive parameters in PyWOFOST crop growth model and making the crop growth model more suitable for the demand of localization yield estimation.a parameter calibration method of’pixel optimization calibration and pixel assimilation correction’was proposed and implemented.The four sensitivity parameters of LAImax were optimized by the time series observed LAI and Particle Swarm Optimization(PSO)method to improve the localization parameters.The localization parameters were inputted into the PyWOFOST crop growth model and using the EnKF algorithm to assimilate the temporal observation variables and crop growth model to adjust the four sensitive parameters of TWSO.Based on this method,the yield estimation accuracy of assimilating LAI,SM,LAI and SM variables at single point scale were analyzed when different meteorological data were used.The results showed that the single-point yield estimation accuracy was the highest with assimilating LAI variable alone when the SM inversion accuracy was low and the site meteorological data was used.And the R~2,RMSE,MAE,Bias of assimilating LAI variable alone reached 0.87,468.64kg/ha,385.70kg/ha,103.08kg/ha.The sensitive parameters selected should match the assimilation variables.The yield estimated accuracy using site meteorological data was higher than using NASA Power meteorological data.(4)In order to solve the problem of low computational efficiency of assimilating remote sensing information and crop growth model in regional yield estimation,a rapid yield estimation method of regional winter wheat based on coupling crop growth model and machine learning algorithm was proposed and implemented.The method of‘pixel optimization calibration,pixel assimilation correction’,EnKF algorithm and NASA Power meteorological data were used to assimilate time series LAI inverted by PROSAIL model into PyWOFOST crop growth model to estimate the yield of a certain number of winter wheat at different growth sampling points.The RFR machine learning algorithm model was established by simulating the multi-point yield and the multi-temporal NDVI of the point.And the established RFR model and the regional NDVI were used to estimate regional winter wheat yield.The results showed that the regional yield estimation accuracy of coupling crop growth model and machine learning algorithm reached 95%in the experimental area,and was better than 86%in2020 and 2021 in Hebi City,and the calculation time required for regional yield estimation was also significantly reduced.This dissertation provided a new theoretical and methodological support for the subsequent regional winter wheat yield estimation and realizes the needs of rapid and accurate large-area yield estimation.