Multivariate Random Forest For Predicting The Spatial Distribution Of Soil Particle Size Fractions And Analysis In The Loess Plateau

Soil particle size fractions describe the contents of sand,silt and clay.As a basic soil physical property,soil particle size fractions strongly affect various physicochemical and hydrological properties of soil,further affecting the ecology,agriculture and water resources in the region.Therefore,the spatial distribution of soil particle size fractions has an important role for research in the fields of geology,ecology and agriculture.The predicted results of soil particle size fractions should satisfy that san,silt and clay contents at each location are not less than 0 and add up to 100%.The current common method is to transform the soil particle size fractions by log-ratio transformations before prediction,but it may affect accuracy of prediction results.In this paper,the potential relationship between sand,silt and clay is considered and a multivariate random forest model is proposed for the combined prediction of soil particle size fractions,which is able to obtain satisfying results.In this paper,243 soil samples and spatial environmental data related to remote sensing,topography and climate were collected for predicting the spatial distribution of soil particle size fractions in the Loess Plateau.Five prediction methods were used,including multivariate random forest(MRF),additive log-ratio transformation combined with random forest(ALR-RF),centroid log-ratio transformation combined with random forest(CLR-RF),isometric log-ratio transformation combined with random forest(ILR-RF),and random forest(RF)for independent prediction of soil particle size fractions.The accuracy of these five prediction methods is compared using three different proportions of the training and test sets.The soil particle size prediction results of the five methods were analyzed at a ratio of 4:1 between the training and test sets.Finally,the spatial distribution of soil particle size fractions in the Loess Plateau were generated using MRF,ALR-RF,CLR-RF,ILR-RF and RF,respectively,and the results were compared.Main conclusions:(1)In the results using the three proportions of training and test sets,MRF consistently has better accuracy and stability in predicting sand,silt and clay contents,with R~2,CCC being the largest and RMSE,MAE being the smallest.(2)The top ranked environmental data variables in terms of relative importance obtained from MRF,ALR-RF,CLR-RF,ILR-RF and RF were almost identical,with nighttime land surface temperature(LST＿N)being particularly important for the prediction of soil particle size fractions.(3)The spatial distribution patterns of soil particle size fractions obtained from MRF,ALR-RF,CLR-RF,ILR-RF and RF are more consistent.The spatial variability of soil particle size fractions in the Loess Plateau is strong,with the contents of silt and clay increase from northwest to southeast,and the content of sand decreases from northwest to southeast,and the overall content of silt is much higher than sand and clay.