A machine learning approach to predicting plant available phosphorus that accounts for soil heterogeneity and regional variability

Abstract

Purpose Mehlich-3 extractable P, Al, Ca, and Fe combined with pH can be used to help explain soil chemical processes which regulate P retention, such as the role of Al, Ca, Fe, and pH levels in P fixation and buffering capacity. However, Mehlich-3 is not always the standard test used in agriculture. The objective of this study is to assess the most reliable conversion of Mehlich-3 Al, Ca, Fe, and P and pH into a commonly used soil P test, Morgan’s P, and specifically to predict values into decision support for fertiliser recommendations. Methods A geochemical database of 5631 mineral soil samples which covered the northern area of Ireland was used to model soil test P and P indices using Mehlich-3 data. Results A random forest machine learning algorithm produced an R2 of 0.96 and accurately predicted soil P index from external validation in 90% of samples (with an error range of ± 1 mg L−1). The model accuracy was reduced when predicted Morgan’s P concentration was outside of the sampled area. Conclusions It is recommended that random forest is used to produce Mehlich-3 conversions, especially when data covers large spatial scales with large heterogeneity in soil types and regional variations. To implement conversion models into P testing regimes, it is recommended that representative soil types/geochemical attributes are present in the dataset. Furthermore, completion of a national scale geochemical survey is needed. This will enable accurate predictions of Morgan’s P concentration for a wider range of soils and geographical scale.