Research On The Yield Estimation Method Of Corn By Coupling Of Crop Growth Model And Light Use Efficiency Model

As one of the most important food crops in the world,corn is widely distributed in many countries and regions.As early as the 1970s and 1980s,combining remote sensing images and ground survey data for applied research such as corn planting distribution,leaf area index information(LAI)extraction and yield estimation proved to be an effective means.Timely and accurate acquisition of these information from remote sensing images is of great significance for ensuring my country’s food security,agricultural policy formulation,international and domestic food trade and sustainable development.Although remote sensing data can quickly and widely obtain various information on the earth’s surface,there are still many problems and challenges in the process of obtaining such information.For example,the current medium-to-high resolution crop distribution products for agricultural applications are missing or updated slowly in some areas,which cannot meet the needs of crop growth monitoring and yield estimation.There are often"ill-posed"problems in the retrieval of LAI from remote sensing images.Fusion of crop growth model and remote sensing data can effectively improve the accuracy of crop yield simulation,but the crop growth model has difficulty in obtaining regional parameters,coupled with the low calculation efficiency of the model,it can only be roughly estimated on the regional scale,and it is difficult to estimate it on the plot scale.Although the light use efficiency model is weaker in mechanism than the crop growth model,it has the advantages of less input parameters and fast calculation speed.How to combine the two to give full play to their respective advantages to meet the needs of yield estimation at the plot scale still needs more in-depth research and practice.In view of the above problems,this paper mainly studies how to use machine learning algorithms to obtain key crop information such as corn distribution and leaf area index from remote sensing data.Based on these information,the feasibility of coupling crop growth model and light energy use efficiency model in corn yield estimation was discussed,and the following conclusions were mainly drawn:(1)Compared with machine learning algorithms such as decision tree(DT),random forest(RF),and support vector machine(SVM),the deep neural network model(DNN)has higher accuracy in the classification of time series remote sensing images of different lengths,so it is selected the DNN model extracted the corn classification information in the study area.First of all,this paper proposes a method that combines historical crop distribution data and ensemble learning to generate a training data set with high precision.Based on this data set,the classification performance of various machine learning algorithms is compared and evaluated.The research results show that the classification accuracy of DNN is higher than that of other algorithms,whether it is for single-temporal or multi-temporal remote sensing images.In the whole growth period of crops,July and August are the best time for distinguishing crop types such as corn,and the kappa coefficients are 0.78 and 0.79 respectively.Using multi-temporal time series can significantly improve crop classification accuracy compared to single-temporal time series.The kappa coefficient increases with the length of the time series,and reaches 0.94 as early as August,and reaches a maximum of 0.95 during the entire growth period.Compared with the verification points collected in the field and the existing crop classification products,the classification results of rice and corn extracted have a high consistency.This shows that the sample data set and DNN model generated by the method in this paper can meet the needs of corn classification information extraction.(2)An ensemble learning method can improve the inversion accuracy of leaf area index(LAI)of a deep neural network model(DNN)for various crop types including corn.This paper proposes a two-stage ensemble learning method,which mainly improves the accuracy of DNN model inversion LAI from two aspects of training data distribution and multiple model integration,and selects three crop types of corn,soybean and sunflower to measure LAI on the ground to validate the effectiveness of the method in this paper.The research results show that when the DNN model is used for LAI inversion,the model itself has certain uncertainties.The risk of uncertainty brought by a single model to the LAI inversion results can be reduced by integrating multiple models.Compared with the uniform distribution and normal distribution assumptions commonly used in existing research and the LAI results retrieved by S2Toolbox of the SNAP software,the LAI retrieval accuracy of the three crop types has been improved by the method in this paper.The overall RMSE of the LAI inverted by the method in this paper is 0.488)~2?8)~2,which is lower than the 0.748)~2?8)~2,0.668)~2?8)~2 and 0.718)~2?8)~2 of the two hypothetical distributions and S2Toolbox respectively,and the inversion results and observations have a high correlation.In addition,the research results also found that,in addition to the influence of data distribution,the use of prior knowledge of crop types can reduce the interference between types when performing medium-to-high resolution LAI inversion,that is,designing data sets for each crop type separately can improve the accuracy of LAI inversion.(3)The coupled crop growth model(WOFOST)and light energy use efficiency model(CASA)can be used to estimate corn yield,and the feasibility and effectiveness of the method are verified by the measured yield.First,methods to improve the accuracy of WOFOST in simulating the corn growth process were studied.Based on the analysis of the sensitive parameters of the crop growth model,and using the inverted corn LAI data as a reference,the DREAM algorithm with a high optimization speed was introduced to calibrate the main sensitive parameters of the WOFOST model.Using corn LAI as the assimilation variable of WOFOST model,the dry matter accumulation and consumption during the growth and development of corn were simulated.The results show that compared with the default parameter values,the correlation between the simulated yield value and the measured value of corn after calibration and data assimilation has improved,and the correlation coefficient has increased from 0.22 to 0.44 and 0.58 respectively.The above studies show that combining parameter calibration and data assimilation methods can improve the accuracy of WOFOST model in simulating the growth and development process of corn.Then,the daily gross primary productivity(GPP)of corn was calculated using the improved CASA model.Due to the time discontinuity of medium and high-resolution satellites,a method of establishing a lookup table based on the Logistic function is proposed on the basis of analyzing the trend of GPP over time.Using this method,the cumulative GPP in the whole growth period of corn was obtained.The research results show that the cumulative GPP calculated by the method in this paper has a strong correlation with the yield,and the simulated results are lower than the measured values,which may be related to the CASA model adopting a strong stress factor calculation method.Finally,the final yield was calculated by combining cumulative GPP and corn dry matter accumulation and distribution simulated by WOFOST model,realizing the loose coupling of the two.After coupling the two,during the corn growth period,in the area where the images are acquired over four time periods,the correlation coefficient between the simulation results and the measured yield can reach up to 0.8,and the RMSE is 0.99 tons/ha.Compared with the simulation results of the crop growth model,the yield estimation accuracy has been greatly improved.The above research results show that the combination of crop growth model and light energy use efficiency model can meet the application requirements of corn yield estimation in medium-to-high resolution remote sensing images.