Study On Sparse Sample Lycium Barbarum Yield Prediction Model Based On Deep Learning

In statistics,the regression problem is a statistical process by learning a set of functions to estimate the relationship between dependent variables and independent variables.The process of finding this set of functions will be very difficult when the independent variable data that can be used for statistics can not effectively express the complete mapping relationship.There is a statistical mapping relationship between Lycium barbarum yield and climate factors,but this mapping relationship is affected by other potential factors and incomplete data statistics,so it is difficult to obtain an effective regression model that can be used to predict yield in the following year.As a conditional probability distribution model that can be used for regression prediction of finite dependent variables,probability regression can effectively solve the problem of incomplete statistical data caused by sparse samples and difficult to obtain effective prediction results.However,for multi-modal sparse samples,it is difficult to capture more complex conditional probability distributions with Gaussian process regression as a typical probabilistic regression model.Although multi-modal Gaussian process can effectively solve this problem,its expensive calculation time and computational space cost are difficult to ignore.Generating adversarial networks can also be used to learn complex conditional distributions,and various derivative models can effectively capture learning high dimensional complex distributions,so it is expected to be an effective method to solve multi-modal sparse samples.By combining the principle of probability regression with the generative adversarial network model,this thesis puts forward a probabilistic regression prediction model design based on conditional adversarial generative network.With Lycium barbarum data and analog data to carry out regression prediction,the difference between mainstream probabilistic regression model and conditional adversarial generative network could be reviewed.The main research contents and innovation points of this thesis can be divided into the following three parts:1.Based on 10 years of output,rainfall,sunshine and temperature of Lycium barbarum in Ningxia region,the basic experimental samples were constructed to study the regression model of yield prediction.On the basis of studying the influence of climatic conditions on the growth and development of Lycium barbarum,a linear and nonlinear regression model between climatic conditions and Lycium barbarum yield was constructed and predicted by regression.Meanwhile,a deep learning platform based on Docker service and remote SSH link is built,which uses TensorFlow2.0 and Jupyter Notebook as the infrastructure for deep learning training.2.Study GPR defects in multi-modal sparse sample regression prediction.In this thesis,three simple nonlinear simulation data and one multi-modal nonlinear simulation data are constructed as experimental samples of simulation data.The sparse samples are made by reducing the sample size of the data and controlling the range of characteristic points.with GPR as the benchmark of the probabilistic regression model,radial basis functions(Radial Basis Function,RBF)and Matern 32 as kernel functions are used to compare the regression prediction defects of GPR under multi-modal sparse samples,respectively.3.Study CGAN advantages in multi-modal sparse sample regression prediction.On the basis of studying GPR multi-modal sparse sample regression to predict defects,this thesis simplifies the method of approximate conditional probability distribution by constructing a CGAN based probability regression model.With the same simulated data samples,the comparison analysis includes the difference of regression prediction results caused by noise,model and loss function.Finally,through the comparative study,the optimal model construction method is obtained,and the Lycium barbarum data is taken as the experimental sample to compare the regression prediction results with the GPR.The study found that CGAN performed better than GPR in Lycium barbarum data and multi-modal sparse samples.