Application Study On Soil Organic Carbon Stock Estimation Based On Improved Temporal Convolution Network

Forest soil organic carbon(SOC)is an important component of forest ecosystems,which can maintain the balance and stability of ecosystems,and is also important for studying the carbon sink function and carbon cycle mechanism of forest ecosystems.At present,the modeling effect of machine learning and deep learning in predicting soil organic carbon(SOC)is not satisfactory,which leads to considerable errors in estimating soil organic carbon stocks.In this thesis,a hyperspectral estimation model of soil organic carbon stock based on Self Attention Temporal Convolutional Network(SATCN)combined with soil type method is proposed for a total of 206 soil samples collected from state-owned Huang Coronation Forestry and state-owned Yachang Forestry in Guangxi.The main research contents and conclusions are as follows:(1)Determination of soil spectral pre-processing methodIn order to explore the hidden information in soil spectral data,four preprocessing methods,SG smoothing and noise reduction as well as first-order differentiation,second-order differentiation,standard normal transformation and multiple scattering correction,were used.The modeling effects of three models,LSTM,PLSR and SVM,were compared and analyzed under different preprocessing methods.The results show that the R2 of the three models of LSTM,PLSR and SVM are improved by 6.3%～13.8%,9%～27.5% and 10.5%～23.6%,respectively,by using SG smoothing followed by the first-order differential transformation.This indicates that first-order differentiation is a better spectral preprocessing method.(2)Establishment of a hyperspectral prediction model for soil organic carbon content based on improved time-convolutional networkTo address the problems of poor modeling and low prediction accuracy of the prediction model under small sample data sets,this paper adopts a shallow network structure on the temporal convolutional network(TCN)architecture.A self-attentive layer is introduced into the TCN residual structure to improve the feature learning capability of the model.In addition,L2 regularization is added to the weights of each convolutional kernel to avoid overfitting.In this paper,the first-order differential preprocessing method is chosen and four models,Res Net-13,VGGNet-7,TCN and SATCN,are built.The modeling effects of these four models are compared and analyzed,as well as the effect of SATCN model at different network depths.The results show that the SATCN model can enhance the important features of spectral sequences and improve the feature learning ability and prediction accuracy of the model.Compared with other models,the SATCN model has higher accuracy and excellent model estimation ability.The R2 on the validation set was 0.943,the RMSE was 3.042 g/kg,and the RPD was 4.273.In addition,the modeling effect of the shallow SATCN model was better than that of the deep model.(3)Estimation of soil organic carbon stock based on soil type methodThe estimated organic carbon content obtained from the SATCN model,combined with the soil capacitance,soil thickness and gravel content,the organic carbon densities of Huang Coronation Forestry and Yachang Forestry were 0.069kg/m2 and 0.148 kg/m2,respectively.Therefore,the soil organic carbon stock in the study area of this paper is 2154.6t/hm2.