Research On Prediction Models Of Soil Properties Using Visible/Near Infrared Spectroscopy

The acquisition of soil composition information is the basis for developing Formula Fertilization by Soil Testing projects and studying the law of crop growth,which plays an important role in the effective management and utilization of land resources and the precision operation of the planting industry.Near infrared spectroscopy analysis is a technique for identifying material varieties or quantitative analysis of chemical components by using the spectral characteristics of material.It has the advantages of non-contact measurement,simultaneous prediction of multi components and low cost.The application of near infrared spectroscopy in the acquisition of soil composition information can greatly reduce the cost of soil information acquisition and promote the improvement of Formula Fertilization by Soil Testing and precision agriculture.The key to obtain soil composition information by near infrared spectroscopy is to establish the relationship between soil spectra and soil composition.Soil samples collected from a large area are highly diverse.In this paper,the calibration models of soil composition based on near infrared spectroscopy are mainly studied,especially the universal calibration models can be used in a large area.The performance of linear models and influencing factors are studied,and two new nonlinear models based on neural networks are proposed.The main contents and conclusions are as follows:1.Soil organic carbon prediction based on linear models.The multiple linear regression model,the principal component regression model,the partial least square regression model and the stepwise regression model are four common linear calibration models.In this paper,four linear models mentioned above are applied to the prediction of soil samples collected from a large area and the effects on prediction results using different spectral preprocessing methods are studied.In this paper,the choice of input includes the original and derivative spectra,the effects of different sampling intervals and different numbers of calibration samples are discussed.2.Soil organic matter grade prediction based on improved auto-encoder.Improved auto-encoder model is optimized based on traditional auto-encoder model.The low dimensional feature representation of high-dimensional spectra extracted by the improved auto-encoder model can not only reconstruct the input,but also predict the soil composition.In this paper,improved auto-encoder model is applied to the grade classification according to soil organic matter content,and the accuracy of prediction is acceptable.3.Soil organic carbon prediction based on convolutional neural networks model.In this paper,the convolutional neural networks model is used to calibrate the model of near infrared spectroscopy.The model uses the original spectra as input,component information as output.Five convolutional neural networks models with different depths are designed,and the effect of depth on model prediction is studied.Many soil samples are used to train the calibration model,and the complex spectral features can be extracted automatically.The experimental results show that the nonlinear models can be used to establish the calibration models in a large scale and provide better prediction than linear models.Both improved auto-encoder model and deep convolution neural network model proposed in this paper can be used as calibration models for soil samples collected in a large area.