ABSTRACT A comprehensive understanding of the spatial distribution of soil organic matter (SOM) is essential for conserving soil fertility, ensuring food security and supporting sustainable agricultural management. Among digital soil mapping (DSM) techniques, geostatistical methods and machine learning (ML) models are the two most widely used approaches, each offering distinct advantages and limitations. Random Forest (RF), a representative ML model, effectively captures complex non‐linear relationships between soil properties and environmental covariates but typically overlooks spatial dependencies. In contrast, the Integrated Nested Laplace Approximation with Stochastic Partial Differential Equation (INLA‐SPDE) is a geostatistical method that explicitly accounts for spatial structure. To harness the complementary strengths of both approaches, this study integrates RF and INLA‐SPDE for SOM mapping in the cropland of Guangzhou City, a region characterized by lateritic red soils. Four variable selection strategies, including variance inflation factor with stepwise regression based on the Akaike information criterion (VIF+StepAIC), recursive feature elimination (RFE), forward recursive feature selection (FRFS) and Boruta, were evaluated to identify the optimal modelling strategy. Results showed that RF ( R 2 = 0.40–0.73) outperformed INLA‐SPDE ( R 2 = 0.19–0.42) across all variable selection methods, highlighting the importance of modelling non‐linear relationships for spatial prediction. Notably, coupling RF with INLA‐SPDE predictions led to a maximum accuracy improvement of 86.5% (based on VIF+StepAIC), demonstrating that incorporating spatial information into the RF model significantly enhances predictive performance. These findings underscore the potential of integrating geostatistics and machine learning for improved SOM mapping in DSM applications.
Zhuo et al. (Wed,) studied this question.