Prediction Model Of Soil Available Potassium Content Based On Convolutional Neural Network

The technology of how to obtain soil composition information quickly and accurately is the basis for carrying out soil formulation,precise fertilization and research on plant growth cycle,and is important for effective statistics of land resources and refinement of forestry production.Hyperspectral analysis is a technique to classify items or quantitatively analyze the chemical composition of substances by analyzing the spectral characteristics data of substances and mining the deep information of the data.The advantages of hyperspectral analysis are mainly reflected in low cost,rapid measurement and long-distance contactless prediction.The use of hyperspectral analysis technology in the process of rapid acquisition of soil composition information can reduce the cost of information acquisition and has positive significance in soil formulation for accurate fertilization and precision forestry.The use of hyperspectral analysis to solve the problem of quantitative analysis of soil nutrients focuses on the construction of the relationship between hyperspectral data and soil composition content.In this dissertation,statistical learning methods are applied to the construction of soil fast-acting potassium content prediction models for soil hyperspectral sample data of eucalyptus plantations in state-owned Huangcong forestry in Guangxi Zhuang Autonomous Region.The effects of different mathematical transformation methods on the prediction accuracy of linear regression models are investigated,and three convolutional neural network models with different network structures are constructed to train the soil hyperspectral data,and different loss functions,different step sizes and different convolutional kernels on the prediction accuracy of convolutional neural network models.The main research contents and conclusions are as follows.(1)linear model-based prediction of soil fast-acting potassium content.In this paper,the original hyperspectral data were preprocessed using the moving tie filter algorithm to eliminate the hyperspectral bands with small variations,and the preprocessed data were mathematically transformed using the first-order derivative,the second-order derivative,and the logarithm of the inverse,and the Pearson correlation coefficients were used to determine the correlation between the soil hyperspectral reflectance and the fast-acting potassium content in the soil after different mathematical transformations,and the forty spectral bands with the largest absolute values were used as input data for the regression prediction model,and the linear regression model was constructed using the principal component analysis algorithm as well as the Lasso variable selection algorithm,and the prediction effects of using the two linear regression models on the hyperspectral sample data of eucalyptus soil in Huang Coronation Forestry were compared,and in a total of eight sets of experiments with the two models and four mathematical transformation methods,analysis of the effect of hyperspectral reflectance data on linear model prediction results for each model using different mathematical transformation methods,the comparison found that: the best prediction effect using the PearsonFDR-Lasso variable selection model,and the prediction effect of this model was verified using the test data set with a coefficient of determination of 0.8126 and a root mean square error of 6.6756,revealing the influence of different spectral preprocessing methods and different models on the prediction accuracy.(2)Prediction of soil fast-acting potassium content based on convolutional neural network model.In this paper,we use convolutional neural network to process onedimensional soil hyperspectral data and construct a model for predicting soil fastacting potassium content,and improve the loss function of convolutional neural network,taking the processed soil hyperspectral data as input and soil fast-acting potassium content as output.Three different structures of convolutional neural networks were designed for accurate prediction of soil fast-acting potassium content,and the convolutional neural network-based prediction model was trained using eucalyptus soil hyperspectral samples from Huang Coronation Forestry,and the prediction effect of the convolutional neural network model with different loss functions was studied.The prediction effect of convolutional neural network with improved loss function and the Pearson-FDR-Lasso variable selection model was found in the experiments: the improved convolutional neural network model had the best prediction accuracy for soil fast potassium with a coefficient of determination of0.8718 and root mean square error of 5.8774 for the test set.Automatic extraction of hyperspectral feature bands by convolutional neural network models.It was demonstrated that the linear regression model performed relatively stable in the test set and validation set compared with the convolutional neural network in the prediction model of fast-acting potassium content for hyperspectral samples of eucalyptus soils in Huang Coronation Forestry,but the best prediction accuracy was achieved using the improved CNN-3 convolutional neural network model.