Predictive Mapping of Soil Organic Matter at a Regional Scale Using Local Topographic Variables: A Comparison of Different Polynomial Models

Abstract

Borrowing the idea of software engineering, this paper aimed to evaluate the mapping accuracy of soil organic matter (SOM) content from the “black box” perspective by combing regression kriging (RK) with local terrain attributes calculated by different polynomial models. When calculating local terrain attributes, we applied two neighborhood shapes (square and circular) and six frequently used algorithms (Evans-Young, Horn, Zevenbergen–Thorne, Shary, Shi, and Florinsky). Overall, 35 combinations of first- and second-order derivatives were produced as secondary information for RK. For comparison, the ordinary kriging (OK), ordinary cokriging (COK), and universal kriging (UK) were also utilized to map the SOM spatial distribution. The results of the study showed that the RK application outperforms OK, COK, and UK in improving the prediction quality of SOM content in a region where the soil properties were strongly influenced by the toposequence and the altitude was with a wide range. The most accurate mapping result was obtained by the combination of the Evans-Young algorithm and Zevenbergen–Thorne algorithm for the calculation of first- and second-order derivatives, respectively. The mapping results from the higher-order approach (Zevenbergen–Thorne and Florinsky) yielded less prediction errors and the circular-neighborhood method could enhance some algorithms for the calculation of local terrain attributes.