From buffer-based tables to dynamic response curves for soil-specific lime requirement estimation using machine learning
Abstract
Accurate estimation of lime requirement (LR) remains a major challenge in acid soils, as conventional buffer-based approaches, particularly the Shoemaker–McLean–Pratt (SMP) method, rely on a single equilibrium measurement and predefined target pH values, limiting their ability to capture the diversity of soil neutralization responses. This study evaluated whether machine-learning (ML) models can reconstruct incubation-derived neutralization curves and provide flexible, soil-specific LR estimates without reliance on SMP buffer chemistry. A total of 400 soil samples from Eastern North America were incubated at six CaCO₃ application rates. Soil responses were expressed as pH change (ΔpH) and fitted using a Mitscherlich model to derive reference neutralization curves and incubation-based LR. ML models were developed using four predictor classes: mid-infrared (MIR) spectroscopy, visible–near infrared (VisNIR) spectroscopy, routine chemistry (RC), and detailed pedological characterization (DPC), as well as hybrid combinations. ΔpH predictions were robust across all models, with DPC achieving the highest performance (R² = 0.92, RMSE = 0.24 pH units under cross-validation), followed by RC and MIR, while VisNIR and minimal SMP-based models showed lower accuracy. LR estimates derived from reconstructed ML curves consistently exhibited lower prediction errors than regional SMP-based recommendations across all target pH levels. Notably, inclusion of SMP buffer pH did not improve model performance when broader chemical or spectral descriptors were available. These results demonstrate that curve-based ML approaches provide a viable alternative to equilibrium-based buffer methods by enabling continuous, soil-specific representation of liming response. This framework supports a transition from static recommendation systems to dynamic, data-driven strategies, laying the foundation for precision liming.
Related
- A response surface framework for persistence-aware lime requirement
- Dynamic response and damage mechanism of deep overburden layer slope based on shaking table tests
- Shape matters: DEM investigation of geometry-controlled mechanical response in irregular rock fragments under static and dynamic loading
- Dynamic response and energy dissipation-transfer characteristics of coal subjected to impact loading under coupled thermo-gas-stress conditions
- Black-Scholes for the land: modelling savannah response to rain as a dynamic world model
BibTeX
@misc{jouichat2026from,
title = {{From buffer-based tables to dynamic response curves for soil-specific lime requirement estimation using machine learning}},
author = {Hamza Jouichat and Lotfi Khiari},
howpublished = {EarthArXiv},
year = {2026},
month = sep,
doi = {10.31223/x51r56},
url = {https://eartharxiv.org/repository/view/15129/},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - From buffer-based tables to dynamic response curves for soil-specific lime requirement estimation using machine learning
AU - Hamza Jouichat
AU - Lotfi Khiari
PY - 2026
DA - 2026/09/21
PB - EarthArXiv
DO - 10.31223/x51r56
UR - https://eartharxiv.org/repository/view/15129/
AB - Accurate estimation of lime requirement (LR) remains a major challenge in acid soils, as conventional buffer-based approaches, particularly the Shoemaker–McLean–Pratt (SMP) method, rely on a single equilibrium measurement and predefined target pH values, limiting their ability to capture the diversity of soil neutralization responses. This study evaluated whether machine-learning (ML) models can reconstruct incubation-derived neutralization curves and provide flexible, soil-specific LR estimates without reliance on SMP buffer chemistry. A total of 400 soil samples from Eastern North America were incubated at six CaCO₃ application rates. Soil responses were expressed as pH change (ΔpH) and fitted using a Mitscherlich model to derive reference neutralization curves and incubation-based LR. ML models were developed using four predictor classes: mid-infrared (MIR) spectroscopy, visible–near infrared (VisNIR) spectroscopy, routine chemistry (RC), and detailed pedological characterization (DPC), as well as hybrid combinations. ΔpH predictions were robust across all models, with DPC achieving the highest performance (R² = 0.92, RMSE = 0.24 pH units under cross-validation), followed by RC and MIR, while VisNIR and minimal SMP-based models showed lower accuracy. LR estimates derived from reconstructed ML curves consistently exhibited lower prediction errors than regional SMP-based recommendations across all target pH levels. Notably, inclusion of SMP buffer pH did not improve model performance when broader chemical or spectral descriptors were available. These results demonstrate that curve-based ML approaches provide a viable alternative to equilibrium-based buffer methods by enabling continuous, soil-specific representation of liming response. This framework supports a transition from static recommendation systems to dynamic, data-driven strategies, laying the foundation for precision liming.
ER -