Study On Yield Estimation Methods Of American Soybean Under Different Spatial Scales

Beneficial to the people’s livelihood is a novel coronavirus pneumonia.The accurate prediction of crop yields is a real issue related to the national economy and people’s livelihood.Especially in the background of the current severe global new crown pneumonia epidemic situation and the increasing impact of climate change,it is increasingly important to accurately provide crop yield information.Soybean is not only an important grain and oil crop in China,but also one of the most important imported grains in international trade.The United States is the main source of soybean imports in China.Therefore,timely and accurate understanding of U.S.soybean production plays an extremely important role in China’s soybean import and export trade and national economic development.Taking American soybean as the research object,facing the state and county-level scales,based on meteorological model,remote sensing model and APSIM model based on Assimilation Algorithm,this paper establishes soybean yield estimation model by using the methods of multiple regression analysis,stepwise regression analysis and neural network,compares and analyzes the yield estimation results of each model,and selects the model with the best prediction accuracy on different spatial scales.The main conclusions are as follows:（1）In the meteorological yield estimation model,ten day average temperature and ten day precipitation are selected as meteorological factors to establish multiple regression yield estimation model,stepwise regression yield estimation model and neural network yield estimation model respectively.The results show that at the state scale,the prediction accuracy of multiple regression model is the highest 98.87% and the lowest 84.28%;The prediction accuracy of stepwise regression yield estimation model is the highest 97.67% and the lowest89.21%;The prediction accuracy of neural network production estimation model is the highest 97.88% and the lowest 90.54%.At the county level,the prediction accuracy of multiple regression yield estimation model is the highest 98.33% and the lowest 90.33%;The prediction accuracy of stepwise regression model is 97.39% and 87.64% respectively;The prediction accuracy of neural network production estimation model is the highest 97.23% and the lowest 89.99%.In terms of the accuracy of prediction results,the meteorological yield estimation model will be affected by bad weather conditions,which makes the prediction accuracy unstable.（2）In the remote sensing yield estimation model,MODIS-NDVI data and Landsat NDVI data are selected as the analysis parameters of state scale and county scale,and multiple regression yield estimation model,stepwise regression yield estimation model and neural network yield estimation model are established respectively.The results show that the average relative errors of the three yield estimation models are 6.91%,9.06% and 4.90%respectively;The average relative errors at the county level are 6.26%,7.73% and 4.01%respectively.The results show that the prediction effect of remote sensing yield estimation model is not ideal,and the yield estimation accuracy needs to be further improved.（3）In the yield estimation model based on Assimilation Algorithm,the particle filter algorithm is used to assimilate the remote sensing Lai（leaf area index）and APSIM（the agricultural production systems simulator）model to simulate the Lai,the assimilated Lai is obtained,and the multiple regression yield estimation model,stepwise regression yield estimation model and neural network yield estimation model are established respectively.The results show that the average relative errors of the three yield estimation models are 4.34%,6.46% and 2.76% respectively;The average relative errors at the county level are 3.93%,3.27% and 1.30% respectively.Among them,the prediction accuracy of neural network production estimation model at the state scale is the highest 99.28% and the lowest 95.46%;At the county level,the highest prediction accuracy is 99.70% and the lowest is 97.41%.It shows that the yield estimation effect of this model is the best among all yield estimation models,and its stability is also the best,which can meet the needs of soybean yield estimation service to the greatest extent.