Research On Forecasting Method Of Winter Wheat Yield Based On Hybrid Neural Network

Winter wheat is one of the main food crops in my country,and its planting area accounts for about 20% of the total area of the main crops,and its output accounts for19% of the total output.Therefore,it is of great significance for crop production management and food supply in my country to accurately predict the yield of my country’s main winter wheat producing areas before harvesting.However,the growth process of crops is very complex,and the plant morphology of crops exhibits obvious phenological changes in different growth stages,and its yield is non-linearly related to many environmental factors.How to construct a suitable crop feature mining algorithm and make full use of feature information of different dimensions is an important prerequisite for accurate crop yield prediction.Therefore,this paper uses MODIS remote sensing images as the data source,and proposes a CNN-GRU production estimation model based on multi-temporal remote sensing images and a CTrm-SVR production estimation model based on multi-source data fusion,which solves the traditional regression model.The problem of insufficient neural network feature extraction has achieved rapid and accurate yield prediction for the main winter wheat producing areas in my country.The main work is as follows:(1)It is difficult for a single neural network to effectively mine the growth characteristics of crops from remote sensing images.Therefore,this paper proposes an end-to-end fusion neural network model(CNN-GRU),which takes multi-temporal remote sensing images as model input,uses CNN to extract deep spectral-spatial features from the images,and converts them into long-term sequences On this basis,GRU mines the time dependence of winter wheat throughout the growth period,and finally uses the output of the fully connected layer of a single neuron to predict yield.The results show that the root mean square error and the mean absolute error of the CNN-GRU yield estimation model are 818 and 560 kg/hm2,respectively,which are better than other comparative models,and the prediction accuracy of this model in the middle and late growth period of winter wheat is comparable,and it can accurately predict Winter wheat yield on a large scale,i.e.the model was able to predict winter wheat yield 2 months before its harvest.(2)When forecasting yield in large-scale regions,the growth periods of crops in different regions are different.To address this dilemma,this paper proposes an attention-based hybrid neural network model(CTrm-SVR).In this model,referring to the encoder structure of Transformer,the self-attention mechanism is used to mine the dependencies of any time series,and the optimal growth period for crop yield prediction is determined in a dynamic weighted manner,making up for the large-scale yield prediction in different regions.Crop phenological differences.In addition,considering that there are many environmental factors affecting crop growth,it is difficult to obtain data,and the crop growth environment in adjacent areas has a strong spatial and temporal correlation.Therefore,this paper uses the geographic location information of the prediction unit to replace a large number of environmental factors,and designs a mutated Laplace kernel function,introduces support vector regression to fuse the deep features extracted by the neural network and the location information of the prediction unit,and replaces the fully connected layer.Final production forecast.The results show that the CTrm-SVR yield estimation model has better performance than other comparative models,and can effectively solve the problem of crop phenological differences in different regions on a large scale.