Improving spatial prediction of soil organic matter in central Vietnam using Bayesian-enhanced machine learning and environmental covariates

Abstract

Soil organic matter (SOM) has a vital role in maintaining soil quality and ecosystem functions. However, predicting its spatial distribution remains a challenging task since it was affected by various environmental covariates. To address this limitation, a novel approach by integrating Bayesian technique into the random forest (RF) algorithm was proposed in this research. A total of 94 surficial soil samples from the top 30 cm and eight key environmental covariates were utilized for training and testing with a 70:30 ratio. According to the results, the enhanced RF model demonstrated a significant improvement in accuracy (RMSE = 0.31%; MAE = 0.25%, R2 = 0.79, Acc = 0.81) compared to the traditional RF model (RMSE = 0.66%; MAE = 0.48%, R2 = 0.10, Acc = 0.61). The four environmental covariates including rainfall, distance to the sea, distance to water bodies, and altitude explained 74.07%, and 75.37% of the variability in SOM content in traditional and enhanced RF models, respectively. Locations with high SOM content were characterized by abundant rainfall, a greater distance from the sea, proximity to rivers, and low elevations. These findings introduce a reliable approach for predicting the spatial distribution of SOM in the context of complex environmental changes.