Research On Soil Nutrient Detection Method Based On Near-infrared Spectroscopy And Deep Learning

In order to realize the rapid detection of soil nutrient information,a soil nutrient prediction model based on one-dimensional convolutional neural network(CNNR)and visible light near infrared spectroscopy is proposed,focusing on the following two Carry out research.(1)Model optimization.Optimizing hyper parameters such as the dimension of the input model's spectral data,the number of layers of the CNNR model,the size of the convolution kernel,and the size of Batch Size,and finally a four-layer volume with an input dimension of 420 wavelength points and a convolution kernel size of 9 is determined Product neural network;in order to eliminate the baseline shift and other background effects in the spectrum,the spectrum is preprocessed by smoothing,first-order derivation,standard normal variable transformation,multiple scattering correction,standard normal variable transformation and detrending,And the results show that the model has the best prediction effect by using the soil spectral data after firstorder derivation processing.The root mean square error of the model for soil p H,organic carbon,calcium carbonate,total nitrogen,phosphorus,and exchangeable potassium are respectively0.3549,7.9951,23.6800,0.5151,22.4431,119.7596.Later,the CNNR model was compared and analyzed with the visual geometry group network based on two-dimensional convolution and the traditional model back propagation neural network and partial least square regression.The results showed that the generalization ability of the CNNR model was much greater than that of the partial least square regression model.Increasingly,p H increased by 14.28%,organic carbon increased by23.28%,calcium carbonate increased by 19.89%,total nitrogen increased by 34.98%,phosphorus increased by 74.47%,and exchangeable potassium increased by 123.70%.(2)Multi-task learning model creation.Multi-task Convolutional Neural Network Regression(MTL?CNNR)based on hard parameter sharing is used to achieve simultaneous detection of soil p H,organic carbon,calcium carbonate,total nitrogen,phosphorus,and exchangeable potassium,and the results Compared with the CNNR model for single soil nutrient detection,the results show that the detection performance of the MTL?CNNR model is better than that of the CNNR model.The MTL?CNNR model has similar goodness of fit for soil p H,organic carbon,calcium carbonate,total nitrogen,phosphorus,and exchangeable potassium.Compared with the CNNR model,it is increased by-0.46%,3.52%,0.30%,3.04%,21.15%,and 20.24% respectively.After that,I explored the influence of Attention mechanism and preprocessing method fusion on model performance.The results show that Attention mechanism and preprocessing method fusion can improve the generalization performance of the model.Among them,the overall performance of the MTL?CNNR model based on feature fusion is the best,and the model has the best effect on soil nutrients.The coefficients of determination are p H: 0.9211,organic carbon: 0.8280,calcium carbonate: 0.9672,total nitrogen: 0.8473,phosphorus: 0.5224,exchangeable potassium: 0.6339.This study explored the effects of two models of single-task convolutional neural network and multi-task convolutional neural network on soil nutrient performance detection,and established a single-task convolutional neural network model and multi-task convolution based on soil visible light near infrared spectroscopy.The neural network model can provide a theoretical basis for the development of rapid soil nutrient detection instruments.