Crop Yield Prediction Based On Deep Learning

In recent years,crop yield prediction has become a research hotspot in the field of agricultural science,and has a key role in solving the problem of food production.Therefore accurate and timely prediction of crop yield is of great significance for the development of relevant national food policy,while providing a reasonable basis for agricultural decision-making,but also to provide an important basis for crop improvement measures.The predicted results can be used to observe the impact of crop growth cycles,soil changes,rainwater distribution and other factors on crop yields.Although the traditional crop yield prediction method can obtain an estimated crop output value and reflect the current growth situation of the crop to a certain extent,its prediction accuracy is low,and it differs greatly from the real output value,and it cannot achieve good results.In view of the above problems,after in-depth study of the development and research status of crop yield prediction,comparative analysis of the advantages and problems of domestic and foreign yield prediction algorithms,This paper proposes a deep neural network model based on remote sensing image data fusion to predict the output value of soybeans at the county level in the United States.The research mainly completed the following aspects:(1)Conduct in-depth research on the relevant theoretical knowledge of crop yield prediction,and analyze the applicability and feasibility of the algorithm.Multi-band satellite remote sensing images can display crop information more comprehensively,using remote sensing image feature fusion data as model input data,integrate the characteristics of vegetation health index,global soil moisture data,land surface reflectance and land cover type that best characterize the growth of crops,A kind of data represents a feature.The features after fusion contain more effective information on crop growth,show more rich details of crops,and provide more data support for yield prediction.(2)Aiming at the problems of huge data sets and redundant information,the histogram calculation method is used to reduce the complexity of the model,and the pixels of the remote sensing image are represented by a three-dimensional matrix and divided into different regions for statistics.The required features are extracted through data dimensionality reduction,and redundant geographic location information is eliminated.(3)The crop yield prediction is a typical regression problem,and it has complex time correlation and different information lengths.Based on these two characteristics,CNN network and LSTM network are selected to predict crop yield by two different model combination methods,effectively solving the overall logical problem between input data information.Finally,the results of the fusion data calculated by the histogram are input as input data to the combined neural network model.The features of the main learning data are trained and the final predicted value is obtained to verify the feasibility and optimization results of the algorithm model.The experimental results show that the fusion of the four kinds of feature data and the comparison of the two kinds of feature data can obtain more information about crop growth and improve the prediction accuracy.On this basis,the prediction error of the LSTM-CNN model is reduced by 2.8% compared to the CNN-LSTM model.The remote sensing data after feature fusion combined with the LSTM-CNN model compares the traditional crop yield prediction model with simple process,low prediction cost and experiment result is more accurate.