Abstract. Soil texture and soil particle size fractions (PSFs) play an increasing role in physical, chemical and hydrological processes. Many previous studies have used machine-learning and log ratio transformation methods for soil texture classification and soil PSFs interpolation to improve the prediction accuracy. However, few reports systematically compared the performance of them in both classification and interpolation. Here, a total of 45 evaluation models generated from five machine-learning models – K-nearest neighbor (KNN), multilayer perceptron neural network (MLP), random forest (RF), support vector machines (SVM), extreme gradient boosting (XGB), combined with original and three log ratio methods – additive log ratio (ALR), centered log ratio (CLR) and isometric log ratio (ILR), were applied to evaluate and compare both of them using 640 soil samples in the Heihe River Basin in China. The results demonstrated that log ratio transformation methods decreased skewness of distributions of soil PSFs data. For soil texture classification, RF and XGB showed better performance with the overall accuracy and kappa coefficients, they were also recommended to evaluate classification capacity of imbalanced data according to the area under the precision-recall curve (AUPRC) analysis. For soil PSFs interpolation, RF delivered the best performance among five machine-learning models with the lowest root mean squared error (RMSE, sand: 15.09 %, silt: 13.86 %, clay: 6.31 %), mean absolute error (MAE, sand: 10.65 %, silt: 9.99 %, clay: 5.00 %), Aitchison distance (AD, 0.84) and standardized residual sum of squares (STRESS, 0.61), and the highest coefficient of determination (R2, sand: 53.28 %, silt: 45.77 %, clay: 53.75 %). STRESS was improved using log ratio methods, especially CLR and ILR. For the comparison of direct and indirect classification, prediction maps were similar on the middle and upper reaches and different on the lower reaches of the HRB. Moreover, indirect classification maps based on log ratio transformed data had more detailed information. There is a pronounced improvement with 21.3 % of kappa coefficient using indirect methods for soil texture classification compared to the direct ones. RF was recommended as the best strategy among these five machine-learning models according to the accuracy evaluation of soil PSFs interpolation and soil texture classification, and ILR was recommended for component-wise machine-learning methods without multivariate treatment considering the constrained nature of compositional data. In addition, XGB was preferred than other models when trade-off of accuracy and time was considered. Our findings can provide a reference for other research of spatial prediction of soil PSFs and texture using machine-learning methods with skewed distribution soil PSFs data in a large area.
Mapping of Source and Target Data for Application to Machine Learning Driven Discovery of IS Usability Problems
