Soil Mapping Research Based On Deep Transfer Learning

Soil is one of the important components of the ecosystem and soil organic carbon pool is a major component of the global carbon pool.Soil organic carbon has an important impact on soil quality,food production security,the ecological cycle of the Earth system,and the reduction of atmospheric carbon dioxide levels.Soil cartography is a discipline that combines soil geography with cartography.Soil mapping is a discipline that uses soil geography and cartography to investigate the quantity and quality of soil resources and thus to grasp the combinations and types of soils and the regularity of their distribution.Accurate and cost-effective mapping of soil properties,such as soil organic carbon(Soil Organic Canbon,SOC),is essential to support policy decisions.Soil properties vary with soil depth,but the environmental covariates typically used in soil mapping to predict soil properties at different depths at the same site have the same value,which may not be sufficient to describe vertical variation in soil properties.Meanwhile,using soil depth only as a covariate in digital soil mapping-based models may not accurately predict SOC at a specific depth.To address this challenge,this paper proposes a novel depth transfer learning method based on a two-dimensional convolutional neural network(2D-CNN)model to improve SOC prediction for each soil depth.The details are as follows:(1)This study proposes a model called deep transfer learning and compare this model with the current popular methods.Firstly,this study applies all depths of data to train the 2D-CNN model,and uses the transfer learning method to fine-tune the already trained model with specific layers of data when predicting specific depths of data.Two independent experiments were conducted to investigate the prediction accuracy by crossvalidation: depth-depth experiments(root mean square(RMSE)of 2.547 % and correlation coefficient()of 0.668)and site-site experiments(RMSE of 2.937 % and of 0.374).Deep transfer models generally have advantages over other machine learning models,including random forest(RF)and two-dimensional convolution(2DCNN).A site-to-site approach is advocated over a depth-to-depth approach when properly evaluating the ability of the model in spatial prediction.Overall,the deep transfer learning approach proposed in this paper makes predictions that characterize the horizontal and vertical variation of soil properties with relatively high accuracy.(2)The influence of the two most relevant covariates on the prediction of soil organic carbon was explored.This paper investigates the role of the two most relevant covariates for SOC in soil organic carbon mapping,namely elevation data(DEM)and Mean Longterm Surface Temperature Jun/Jul,in soil organic carbon mapping.(3)The role of additional covariates in improving soil mapping using machine learning is explored.Soil data are critical in soil mapping.The quality of soil data determines the upper limit of soil mapping.In this study,the effect of additional covariates on the effectiveness of soil mapping was explored in addition to the selection of covariates based on autocorrelation.(4)The effect of freezing a layer of the neural network on the prediction effect after performing soil mapping was investigated.Deep learning is a parameter-sensitive method,and different neural layers of the deep neural network are used again to play different roles in making predictions.In this study,the effect of different neural layers on soil organic carbon prediction was investigated by freezing different neural layers and finetuning the remaining layers during prediction.