Research On Near-infrared Soil Carbon Content Analysis Model Based On Convolutional Neural Network

In the double carbon context of carbon neutral and carbon peaking,the carbon sink function of forest ecosystem is of great concern,and forest soil is an important component of forest ecosystem,and estimating the carbon content of forest soil is the basis for studying the carbon sink capacity of forest soil.Near infrared spectrum（NIRs）technique can use the spectral characteristics of substances to identify the species of substances or measure the chemical composition of substances,which has the advantages of rapid,efficient and non-destructive.Applying NIRs technology to the nondestructive detection of soil carbon content can substantially reduce the cost of soil carbon content information acquisition.In order to realize the rapid prediction of soil carbon content information,this paper takes the forest soil of Dongfang Hong forestry field of Beltling Forestry Experimental Bureau in Xiaoxinganling region as the research object,and combines the NIR spectrum technique with the deep residual network（ResNet）and the lightweight network（Mobile Net）in convolutional neural network（CNN）to realize the rapid prediction of forest soil carbon content.（1）NIR prediction model of soil carbon content based on depth residual network.The dimensionality of the input data,the depth of the neural network and the batch size were preferred,and finally,it was determined that 47×47 2D spectral data were used as the input to the model,the batch size was 15,and an 18-layer deep residual network（ResNet18）was selected for training.To eliminate the noise generated by background,particle inhomogeneity,etc.,a single preprocessing method（SG convolutional smoothing,derivative,etc.）and a combined preprocessing method(first-order derivative(1^stD)+SG convolutional smoothing,standard normal variable transform（SNV）+detrending（DT）,etc.)were used to preprocess the spectral data of soil.Model has the best performance with root mean square error（RMSE）and coefficient of determination（R²）of 7.146 0and 0.820 9,respectively,for its validation set.Comparative analysis of the ResNet18 model with the conventional BP neural network（BPNN）and partial least squares regression（PLSR）models and the convolutional neural network model VGG19 reveals that the R² of the ResNet18 model improved by0.33%,14.73%and 39.90%compared to VGG19,PLSR and BPNN,respectively,and the training time was reduced by 58.56%compared to that of VGG19.（2）NIR prediction model of soil carbon content based on MobileNetV1 convolutional neural network model.The NIR prediction model of soil carbon content based on MobileNetV1 was constructed and compared with ResNet18,PLSR and BPNN for analysis.The results show that the R² of MobileNetV1 model is 3.68%and 26.50%better than PLSR and BPNN,respectively,and 9.64%lower than that of ResNet18 deep residual network model.To take advantage of the short training time of MobileNetV1,we try to increase the sample size of the training set to improve the accuracy of the model,and it is found that the MobileNetV1 model has the best performance when the sample size of the training set is 90%of the total sample size,and its RMSE and R² are 5.490 2 and 0.819 1,respectively.Comparing the effect of the sample size of the training set on the performance of different models,the results show that the MobileNetV1 model performs much better than PLSR and BPNN and slightly better than ResNet18 when the sample size of the training set is 90%of the total sample size,and the training time is 63.2%less than that of ResNet18.It is shown that the ResNet18 model can obtain good prediction accuracy with a validation set R²of 0.820 9 when the input spectral data dimension is 47×47,the preprocessing method is 1^stD+SG convolutional smoothing,and the batch size is 15.When the training set sample size is 90%of the total sample size,the MobileNetV1 model can obtain good prediction accuracy with a validation set R² of 0.819 1.The validation set R² is 0.819 1,and the training time can be greatly reduced.This study can provide a theoretical basis for the rapid detection of forest soil carbon content.