Modeling soil pH dynamics with distinct buffering mechanisms: insights from two purple soils

Abstract

Soil acidification models are useful for evaluating measures to mitigate soil acidification under various agronomic practices. However, the appropriate modeling approaches for simulating the soil acidification process have not been adequately studied across soils with distinct buffering mechanisms. This study evaluated the performance differences between a process-based soil acidification model (VSD+) and four machine learning models, including random forest (RF), support vector machine, extreme gradient boosting and decision tree, in simulating pH dynamics of neutral and acidic soils. Two long-term experimental sites were selected with distinct buffering mechanisms on purple soil as an example for the development, calibration and validation of soil acidification models. Results from the RF importance factor analysis indicated that soil background pH was the primary factor influencing the dynamic changes in purple soil pH, followed by meteorological conditions and agronomic practices. pH was then chosen as an essential input variable to developing machine learning models for simulating soil acidification patterns. Machine learning models achieved higher accuracy in neutral soil than the VSD+ model. The RF model gave the best simulation performance, outperforming other machine learning models at both sites, with the highest R² of 0.70 and 0.47 and the lowest MAE of 0.19 and 0.17 for neutral and acidic soils, respectively. In contrast, the VSD+ model exhibited excellent accuracy with acidic soil (R² = 0.95, RMSE = 0.05 and MAE = 0.02) compared to the other machine learning models (R² = 0.20–0.47, RMSE = 0.15–0.23 and MAE = 0.14–0.20). These findings provide information for selecting the most suitable modeling approach to simulate soil acidification process with distinct buffering mechanisms, supporting informed decision-making for restoring soil health and quality.

Graphical abstract

Keywords

Long-term experiments / machine learning / purple soil / soil acidification / VSD+ model

Highlight

	● Soil pH dynamics were modeled in two purple soils with distinct buffering mechanisms.
	● Soil background pH was the primary factor affecting soil pH changes.
	● Machine learning models were superior for neutral soil.
	● The Very Simple Dynamic Model Plus (VSD+) performed excellently for acidic soil.
	● Random forest modeling gave the best accuracy of four machine learning models tested.

RIGHTS & PERMISSIONS

The Author(s) 2025. Published by Higher Education Press. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0)