| In recent years,with the development of science and technology,agriculture has entered the 4.0 era,and the state has also issued a series of documents to promote the rapid progress of agriculture in the direction of intelligence and precision.However,due to the high cost of agricultural data collection,insufficient decision-making model and low degree of intelligence,the problems of low land output rate,resource utilization rate and labor productivity in the process of agricultural production are caused.Aiming at the problems of high soil sampling cost,difficult to obtain soil nitrogen,phosphorus and potassium content and low long-term prediction accuracy of the model in the decision-making of maize precision fertilization,and the lack of identification model of maize disease occurrence degree in the decision-making of maize pesticide application,a prediction model of soil nitrogen,phosphorus and potassium content based on Improved BP neural network was constructed The combined prediction scheme of small sampling and comprehensive model prediction and the identification model of maize disease occurrence degree based on improved deep learning model.According to the content of nitrogen,phosphorus and potassium in soil and the occurrence degree of maize diseases,the precision fertilization and application decision-making model of maize was established,and the intelligent decision-making system was established.After testing,the prediction accuracy of the improved model is high,which can provide decision-making basis for variable rate fertilization and pesticide application.The specific research work is as follows:1.In order to provide a large number of test data for the research and improvement of this model,the data of soil nitrogen,phosphorus,potassium and maize diseases were obtained by different means.In terms of soil nitrogen,phosphorus and potassium data acquisition,using the developed soil sampling grid automatic division program,the maize plot boundary of No.13village,Gongpeng Town,Yushu City,Jilin Province obtained by GPS was divided into grids and the sampling points were determined.According to the sampling points,soil sampling and testing were carried out,and five years of soil nutrient data were obtained;In terms of maize disease data acquisition,3534 open maize disease data were obtained on plantvillage website,and 276maize disease images were taken in the maize experimental field of Jilin Agricultural University by mobile phone.According to the national maize disease classification standard,the obtained maize disease images were divided into three levels:health,general and serious,and the data were expanded.2.Aiming at the problems of high sampling cost and difficult acquisition of soil nitrogen,phosphorus and potassium content,a soil nutrient content prediction model integrating BP neural network,gray wolf algorithm and reverse learning mechanism is proposed.The model was established by using the data of the first four years of soil available nitrogen,phosphorus and potassium data for five consecutive years,and the model was tested by using the data of the fifth year.The test shows that the prediction accuracy of soil available nitrogen in the fifth year is88.8%,which is 7.71%higher than that of BP neural network model;The prediction accuracy of soil available phosphorus reached 88.28%,which increased by 28.83%;The prediction accuracy of soil available potassium reached 91.21%,which increased by 6.79%.The test shows that the prediction accuracy of soil nitrogen,phosphorus and potassium of this model is greatly improved compared with the original model,and the content of soil nitrogen,phosphorus and potassium can be obtained without sampling and detection.3.In the acquisition of soil nitrogen,phosphorus and potassium content,the long-term prediction accuracy of the model will be reduced,and the detection according to the soil sampling grid will lead to huge workload and increased cost.Therefore,based on the idea of combining time and space,this paper proposes a soil nutrient detection scheme combining a small amount of sampling and comprehensive prediction of the model.By sampling and detecting a few sampling points,the measured data of soil nitrogen,phosphorus and potassium are obtained.The difference between the measured data and the predicted data of the sampling points is used to calculate the spatial impact value on the non sampling points,and then the sum of the predicted data and the spatial impact value of the non sampling points is calculated to obtain the verified predicted value.Through the combination of soil nutrient value(measured value,difference),sampling point selection(average,grid optimization),power of distance weight(1,2,3)and the number of known points(4,6,8,10,12),60 different combination schemes are obtained.After testing,the optimal combination is:difference,grid optimization,the power of distance is 1,and the number of known points is 8.At this time,the prediction accuracy of soil available nitrogen,available phosphorus and available potassium is 91.19%,90.27%and 93.10%,which are further improved than that of the improved BP neural network model.4.Aiming at the problem that there is no corn disease occurrence degree recognition model, an improved deep learning corn disease occurrence degree recognition model is proposed.Firstly,the ratios of training set and verification set are 2:1,3:1,4:1 and 5:1 respectively,and the best ratio is determined to be 3:1.Secondly,the parameters of the model are tested,and the optimal parameter combination of Adam algorithm optimization,relu excitation function,linear attenuation,maximum learning rate of 0.0001 and regular term parameter of 0.01 is obtained.Through analysis and comparison,resnext model is determined as the basic model.Aiming at the problem that the first layer convolution kernel can not well extract the characteristics of small corn disease spots,a method of using three 3*3 convolution kernels to replace the original 7*7 convolution kernel is proposed,and the group number C is increased from 32 to 64.Through the test,for the identification of different corn diseases,the recognition accuracy of the algorithm is 98.69%on the open source data set and 97.46%in the real environment,which is higher than the references and the original model.The recognition accuracy of corn disease degree is 89.31%,which is the highest compared with other classical deep learning models and the original model.Verified by the measured data,the recognition accuracy is 90.22%,which can provide a basis for the decision-making of real-time operation of the applicator.5.Aiming at the problem of low intelligence of maize planting,an intelligent decision-making system for precision fertilization and application of maize was realized by using GIS,Internet of things,deep learning and programming technology.The system includes the basic operation of spatial information,soil nutrient query and fertilization decision-making,diagnosis and treatment of maize diseases,pests and Weeds Based on soil moisture,maize disease identification and pesticide decision-making,prescription map making and so on.Through variable rate fertilization,the average amount of fertilization in No.13 village,Gongpeng Town,Yushu City,Jilin Province is631kg/hm~2,which is 68kg/hm~2 less than traditional fertilization,the average yield is 8313kg/hm~2,and the yield is increased by 813kg/hm~2,which achieves the purpose of reducing investment and increasing income.In terms of variable application,the calculation of application amount has been realized at present,and no demonstration application has been carried out. |
PDF Full Text Request