Research On Comparison Of Estimation Methods Of Leaf Area Index Of Northern Summer Maize

Leaf Area Index(LAI)is an important indicator to measure crop growth and yield.However,traditional methods of obtaining leaf area index are mostly manual measurements,which not only wasted manpower and material resources,but also damaged the leaves of crops and affected crop growth.Manual measurements are not easy to promote in large-scale field management.Therefore,for the problem that the leaf area index was difficult to obtain quickly,accurately and widely.The study based on the leaf area index and plant height of the maize collected in Zhaojun Town,Erdos City,Inner Mongolia Autonomous Region in 2018,combined with the multispectral image of the UAV in the same period,8 vegetation indexes and height with strong LAI correlation with maize were selected as the input variables of the inversion leaf area index.The Gradient Boosting Decision Tree(GBDT)algorithm was used to establish the prediction model among the vegetation index,maize height and leaf area index,The prediction model is compared with the model established by Support Vector Regression(SVR)algorithm and Random Forest(RF)algorithm.The main research contents and results of the thesis are as follows:(1)Estimate variable extraction and sample grouping.Before constructing the machine learning algorithm model,by analyzing the leaf area index related attribute characteristics,it was found that the reflectance of the red and near infrared bands was most closely related to the characteristics of the crop leaves and is different from other features.The maize height also shows the strong correlation with LAI.so extracted band information could be combined into a variety of vegetation indexes,and fused maize height information as common variables for LAI estimation;To ensure the applicability and scientificity of the model,the total training sample was repeatedly divided into three groups,each sample group was divided into 65% training set and 35% verification set by hand-out cross validation;Then,using Pearson correlation coefficient analyed the correlation between 8 vegetation indexes,maize height and leaf area index in sample group 1,2 and 3.The results showed that the correlation coefficient between 8 vegetation indexes and LAI were both above 0.66(P(27)0.01),and the maize height and LAI correlation index were more than 0.71(P(27)0.01).Therefore,NDVI,RVI,SAVI,OSAVI,EVI2,RDVI,MASVI,TVI and maize height can be selected as input variables for the model.(2)Leaf area index estimation method.The leaf area index estimation model was constructed by using support vector regression algorithm,random forest algorithm and gradient boosting decision tree algorithm.The three models performed a comparison of the accuracy of the lateral model of maize height and no-height(only 8 vegetation indexes)for each algorithm,and the longitudinal precision comparison among the three algorithms.The experiment proved that the lateral comparison of each algorithm construction model,the estimation accuracy of the three algorithms model leaf area index based on maize height was higher than the corresponding algorithm based on multispectral 8 vegetation indexes.The longitudinal comparison analysis of the three algorithms with the maize height,the gradient boosting decision tree algorithm in the maize leaf area index estimation method showed better performance than the support vector regression algorithm and the random forest algorithm,and was the algorithm with the highest accuracy of maize LAI estimation in the three algorithms.The algorithm increased the weight of the problem sample each time,and the residual of the previous calculation result was used as a new round of training data,which the algorithm could accurately invert the maize LAI.Therefore,the sample group 2model with the best coefficient of 0.7558 was selected in the GBDT algorithm.The training model was used to verify the inversion of the regional maize LAI for the ninth(August 23)experimental data.The accuracy of the GBDT algorithm model was determined by analyzing the average,maximum and minimum values of each research partition of the inversion and actual values.The research showed that the inversion results were basically consistent with the measured results.