Soil acidification is one of the pressing issues confronting global farmland today. Studies indicate that approximately 40% of farmland soils worldwide are at risk of acidification. In China, the topsoil pH of major croplands has decreased by an average of 0.5 units between 1980 and 2000, largely attributed to excessive chemical fertilizer application and atmospheric acid deposition. Soil acidification not only degrades soil quality but also reduces nutrient availability, increases the release of toxic elements such as aluminum and manganese, inhibits crop growth, and even causes yield losses. To effectively manage soil quality, models simulating pH dynamics have become essential tools. However, different soils exhibit distinct acid buffering mechanisms—for instance, some rely on calcium carbonate to neutralize acidic substances, while others depend on exchangeable base cations. So, which model can more accurately simulate pH dynamic changes in purple soils with varying buffering mechanisms?
Professor Xiaojun Shi and Dr. Xuanjing Chen from the College of Resources and Environment, Southwest University, conducted a study to address this question. They selected purple soils from two long-term experimental sites in Chongqing: one neutral purple soil and one acidic purple soil. The research compared the performance of a process-based model (VSD+) and four machine learning models (Random Forest, Support Vector Machine, Extreme Gradient Boosting, and Decision Tree) in simulating pH dynamics of the two soil types. The relevant findings have been published in Frontiers of Agricultural Science and Engineering (DOI: 10.15302/J-FASE-2025658).
The results revealed that model performance varied with soil buffering mechanisms. For neutral purple soil, machine learning models demonstrated significantly higher simulation accuracy than the VSD+ model. Among them, the Random Forest model performed the best, achieving a coefficient of determination (R²) of 0.70, compared to only 0.37 for the VSD+ model. Conversely, in acidic purple soil, the VSD+ model outperformed all machine learning models, with an R² as high as 0.95. Additionally, the study identified soil background pH as the primary factor influencing pH dynamics, explaining 26.8% of the variation, followed by meteorological conditions and agronomic practices. This finding underscores the importance of soil background pH as an input variable in modeling.
This study systematically compared the applicability of process-based and machine learning models in purple soils with distinct buffering mechanisms, establishing that neutral purple soils are better suited for machine learning models while acidic purple soils yield more accurate results with the VSD+ model. These findings provide a scientific basis for simulating acidification in purple soils with different buffering mechanisms, enabling agricultural managers to more accurately predict soil acidification trends and formulate targeted measures (such as rational fertilization and application of amendments) to mitigate acidification and protect soil health.