Robustness of machine learning algorithms to skewness, spatial dependence, and sample size compared to BME in soil fertility prediction

Abstract This study investigates the robustness of machine learning (ML) algorithms compared to the Bayesian Maximum Entropy (BME) framework for predicting four key soil fertility indicators: organic matter (OM), nitrogen (N), phosphorus (P), and potassium (K) under varying data conditions. Using a simulation-based design, twelve scenarios were constructed by systematically varying sample size, spatial dependence, and distribution skewness. Results show that predictive performance is strongly influenced by data characteristics. On average, BME achieved the highest accuracy across these fertility indicators (R $$^2$$ up to 0.85), outperforming ML models by effectively capturing spatial dependence and uncertainty. Among ML approaches, Random Forest (RF) consistently delivered the best performance (R $$^2$$ up to 0.82), followed by Artificial Neural Networks (ANN) and Support Vector Machines (SVM). ML models demonstrated superior performance in data-rich environments with weak spatial dependence, leveraging their ability to model complex nonlinear relationships. In contrast, BME showed greater robustness under strong spatial dependence and limited sample sizes due to its explicit integration of spatial covariance structures and probabilistic information. Additionally, log-transformation of skewed data improved ML model performance, confirming the sensitivity of data-driven approaches to distributional properties. These findings highlight the complementary strengths of ML and geostatistical methods and provide practical guidelines for selecting optimal modeling strategies in digital soil mapping.