Disentangling the drivers of total and available phosphorus distributions in Swiss soils

Phosphorus (P) is essential for plant growth, but excessive accumulation in soils can pose environmental risks, particularly through losses into water bodies. However, despite large spatial variability in P concentration and availability, no nationwide map of soil P exists for Switzerland. Moreover, knowledge of the relative importance of management-related factors compared to soil chemical and pedoclimatic drivers on the distribution of total and available P remains limited.

This study aims to predict and explain the distribution of total and available P pools in Swiss topsoils. For this purpose, we combined machine-learning approaches (ML) using Random Forests with model interpretation based on Shapley values to identify the main drivers controlling the P distribution. As model input, we used total P data measured at 960 sites of the Geochemical Soil Atlas of Switzerland in combination with predictor variables representing land use, soil properties, as well as environmental and geological conditions. To further investigate P dynamics in agricultural soils, we integrated a dataset from the Swiss Proof of Ecological Performance (PEP) subsidy scheme, which comprises 14 years of soil analyses since 2010. Available P pools were operationally defined using CO₂ saturated water extraction as a proxy for immediately plant available P, and ammonium acetate - EDTA extraction (AAE10) representing a larger pool of exchangeable soil P. This dataset includes approximately 150’000 observations, allowing differentiation between arable land and grassland.

As expected, total P concentrations are significantly higher in arable land and in pastures/grasslands compared with forests and alpine areas. Accordingly, interpretable ML measures, including Shapley values, indicate that land use is the most important predictor, followed by the presence of nutrients such as nitrogen (N), potassium (K), and sulfur (S). In contrast, soil chemical properties (e.g. pH, soil organic carbon) and proxies for pedoclimatic conditions, such as temperature or lithology, are less important for the prediction on a national scale.

Across Switzerland, the lowest available P concentrations are observed in north-western and southern regions, whereas the highest concentrations were measured along the Swiss Plateau and in central and north-eastern Switzerland. While P concentrations extracted with CO₂ saturated water are similar between arable crops and grassland, arable soils exhibit systematically higher AAE10 extractable P. Further work will focus on identifying the main drivers of available P pools and their temporal changes across Switzerland, including data from remote sensing and other monitoring programmes.

By combining spatially resolved geochemical data, interpretable machine learning approaches, and long-term agricultural monitoring data, this study provides a framework for identifying key drivers of the distribution of P pools in Swiss soils, thereby supporting targeted and sustainable nutrient management strategies.