Enhancing Sustainable Agriculture: Machine Learning-Based Soil Health Prediction in Permaculture

Abstract

Soil health is central to sustainable agriculture, but remains challenging to assess in diversified agroecosystems such as permaculture. Soil condition reflects the interaction of physical, chemical, and biological properties, but practical assessment commonly relies on a limited set of informative indicators. In this study, a PCA-weighted Soil Health Index (SHI) was constructed from five surface soil indicators: organic carbon, total nitrogen, microbial biomass (PLFA), bulk density, and gravimetric water content. The first principal component explained 75.30% of the total variance. The analysis used 84 observations collected between 2019 and 2021 from permaculture and conventional farming systems across nine locations in Germany and Luxembourg, encompassing different land use types and two soil depths. Permaculture plots showed higher SHI values overall than conventional plots, with the same trend observed across all nine locations, although land use imbalance limited fully matched comparisons. To avoid circular prediction of the PCA-derived target, the five surface variables used directly to construct the SHI were excluded from the predictive feature set. Machine learning models were evaluated using grouped validation in which entire locations were held out from model training. The best-performing full-profile Ridge model achieved an out-of-fold R2 of 0.710, an MAE of 0.159, and an RMSE of 0.214. Out-of-fold SHAP analysis indicated that magnesium, zinc, soil pH, subsoil bulk density, and copper made the largest model-specific contributions to SHI estimation. These findings demonstrate that PCA-based soil health assessment can distinguish systematic differences between studied farming systems and that a leakage-aware, interpretable modeling framework can provide moderate predictive performance across held-out locations. The results should be interpreted as internal evidence from a small multi-location dataset rather than as externally validated or causal estimates of management effects.