Li Decheng

An Insight into Machine Learning AlgorithmstoMapthe Occurrence of Soil MatticHorizon in the Northeastern Qinghai-Tibetan Plateau

Abstract Soil diagnostic horizons, which each have a set of quantified properties, play a key role in soil classification. However, they are difficult to predict, and few attempts have been made to map their spatial occurrence. We evaluated and compared four machine learning algorithms, namely, the classification and regression tree (CART), random forest (RF), boosted regression trees (BRT), and support vector machine (SVM), to map the occurrence of the soil mattic horizon in the northeastern Qinghai-Tibetan Plateau using readily available ancillary data. The mechanisms of resampling and ensemble techniques significantly improved prediction accuracies (measured based on area under the receiver operator characteristic curve score (AUC)) and produced more stable results for the BRT (AUC of 0.921 ± 0.012, mean ± standard deviation) and RF (0.908 ± 0.013) algorithms compared to the CART algorithm (0.784 ± 0.012), which is the most commonly used machine learning method. Although the SVM algorithm yielded a comparable AUC value (0.906 ± 0.006) to the RF and BRT algorithms, it is sensitive to parameter settings, which are extremely time-consuming. Therefore, we consider it inadequate for occurrence-distribution modeling. Considering the obvious advantages of high prediction accuracy, robustness to parameter settings, the ability to estimate uncertainty in prediction, and easy interpretation of predictor variables, BRT seems to be the most desirable method. These results provide an insight into the use of machine learning algorithms to map the mattic horizon and potentially other soil diagnostic horizons.

Mapping Soil Organic Carbon Using Local Terrain Attributes: A Comparison of Different Polynomial Models

Abstract Local terrain attributes, which are derived directly from the digital elevation model, have been widely applied in digital soil mapping. This study aimed to evaluate the mapping accuracy of soil organic carbon (SOC) concentration in 2 zones of the Heihe River in China, by combining prediction methods with local terrain attributes derived from different polynomial models. The prediction accuracy was used as a benchmark for those who may be more concerned with how accurately the variability of soil properties is modeled in practice, rather than how morphometric variables and their geomorphologic interpretations are understood and calculated. In this study, 2 neighborhood types (square and circular) and 6 representative algorithms (Evans-Young, Horn, Zevenbergen-Thorne, Shary, Shi, and Florinsky algorithms) were applied. In general, 35 combinations of first- and second-order derivatives were produced as candidate predictors for soil mapping using two mapping methods (i.e., kriging with an external drift and geographically weighted regression). The results showed that appropriate local terrain attribute algorithms could better capture the spatial variation of SOC concentration in a region where soil properties are strongly influenced by the topography. Among the different combinations of first- and second-order derivatives used, there was a best combination with a more accurate estimate. For different prediction methods, the relative improvement in the two zones varied between 0.30% and 9.68%. The SOC maps resulting from the higher-order algorithms (Zevenbergen-Thorne and Florinsky) yielded less interpolation errors. Therefore, it was concluded that the performance of predictive methods, which incorporated auxiliary variables, could be improved by attempting different terrain analysis algorithms.

Mapping soil organic carbon stocks of northeastern China using expert knowledge and GIS-based methods

The main aim of this paper was to calculate soil organic carbon stock (SOCS) with consideration of the pedogenetic horizons using expert knowledge and GIS-based methods in northeastern China. A novel prediction process was presented and was referred to as model-then-calculate with respect to the variable thicknesses of soil horizons (MCV). The model-then-calculate with fixed-thickness (MCF), soil profile statistics (SPS), pedological professional knowledge-based (PKB) and vegetation type-based (Veg) methods were carried out for comparison. With respect to the similar pedological information, nine common layers from topsoil to bedrock were grouped in the MCV. Validation results suggested that the MCV method generated better performance than the other methods considered. For the comparison of polygon based approaches, the Veg method generated better accuracy than both SPS and PKB, as limited soil data were incorporated. Additional prediction of the pedogenetic horizons within MCV benefitted the regional SOCS estimation and provided information for future soil classification and understanding of soil functions. The intermediate product, that is, horizon thickness maps were fluctuant enough and reflected many details in space. The linear mixed model indicated that mean annual air temperature (MAAT) was the most important predictor for the SOCS simulation. The minimal residual of the linear mixed models was achieved in the vegetation type-based model, whereas the maximal residual was fitted in the soil type-based model. About 95% of SOCS could be found in Argosols, Cambosols and Isohumosols. The largest SOCS was found in the croplands with vegetation of Triticum aestivum L., Sorghum bicolor (L.) Moench, Glycine max (L.) Merr., Zea mays L. and Setaria italica (L.) P. Beauv.