A similarity-based method for three-dimensional prediction of soil organic matter concentration

- A similarity-based method was integrated with depth function for 3D SOM prediction.
- SOM profile distribution was modeled with a piecewise linear and power function.
- The depth function parameters were interpolated using a similarity-based method.
- Similarity-based prediction is an attractive alternative to the “black box” methods.

This paper presents an approach to predicting three-dimensional (3D) variation of soil organic matter (SOM) concentration by integrating a similarity-based method with depth functions. It was tested in a small hilly landscape. A depth function model was constructed to fit SOM profile distribution using a linear relation in the topsoil and a power function in the subsoil. Then, under the assumption that similar environmental conditions at two sites would lead to the development of similar profile morphologies and thus similar depth function parameters, the similarity-based method was used to spatially interpolate the depth function parameters based on their relationships with environmental variables. With the values of the parameters for every location, a 3D map of SOM distribution was generated. The predicted SOM pattern well reproduced the statistical distribution of the pedon dataset used in this study. The overall mean error (ME) was 0.06 g kg− 1 and ratio of performance to deviation (RPD) was 2.34. We conclude that the proposed approach is effective and accurate for 3D SOM prediction. It overcomes two drawbacks of the frequently used pseudo 3D soil mapping approach: (1) the neglect of vertical soil pattern when performing horizontal soil predictions, and (2) the repeated applications of depth function fittings in the mapping process, both of which may lead to prediction errors. Moreover, the similarity-based method is a transparent and traceable prediction process, allowing for easy interpretation of its results. This is useful for understanding soil-environmental relationships and processes. The method thus is an attractive alternative to the commonly used non-linear “black-box” techniques such as artificial neural networks.