Objective  Traditional Mollic epipedon determination relies on field profile excavation , indoor physicochemical analysis and experts' experience, which is hard work, time consuming and highly subjective. In this study, machine learning and digital image processing techniques were used to quickly determine the Mollic epipedon.

Method Soil profile image information from the National Soil Series Survey and Compilation of Soil Series of China dataset was used to model the relationship between soil color and physicochemical properties, and three machine learning classifiers, namely Random Forest, Support Vector Machine and Gradient Boosting Tree, were used to perform the binary classification of the Mollic epipedon and the non-Mollic epipedon.

Result All three classifiers could discriminate the Mollic epipedon, among which Random Forest has the best classification effect, with an average accuracy of 0.80 and an average kappa coefficient of 0.55. The a, b, u, and v variables in the super-red band 2R-G-B and the La*b*, Lu*v* color models were important classification feature variables for discriminating the Mollic epipedon samples from the non-Mollic epipedon samples, whereas the La*b*, Lu*v* L variables in the color model, and several X, Y, and Z variables in the CIEXYZ color model contribute less to the classification. Compared to the sRGB color model, the La*b*, Lu*v* color model has a better classification effect.

Conclusion The Mollic epipedon discrimination method based on soil profile images proposed in this study can somewhat assist the discrimination of traditional methods to identify the Mollic epipedon quickly and finely, but further research is still needed for soil profiles with complex soil configurations.