Digital Mapping of Soil pH and Electrical Conductivity: A Comparative Analysis of Kriging and Machine Learning Approaches

Abstract

Soil pH and electrical conductivity (EC) are critical soil properties influencing agricultural productivity and environmental sustainability. This study evaluates the performance of stacked machine learning models in predicting and mapping soil pH and EC values. Base models such as Ordinary Kriging (OK), Universal Kriging (UK), and Disjunctive Kriging (DK) were employed, and their outputs were integrated into a Multilayer Perceptron (MLP) neural network meta-model. The results reveal the superior performance of the MLP meta-model across all metrics. For instance, in predicting pH, the MLP model achieved an RMSE of 0.028, an MAE of 0.020, and an R² of 0.858 on the training dataset. For EC predictions, the MLP model outperformed others on the test dataset, with an RMSE of 0.039, an MAE of 0.028, and an R² of 0.900. In contrast, the UK and DK methods exhibited lower accuracy, particularly on test datasets. This study shows the advantage of modern machine learning algorithms in modeling nonlinear spatial relationships and their significant potential in digital soil mapping. The findings demonstrate the applicability of these approaches in enhancing agricultural productivity and supporting sustainable soil management practices.