In precision agriculture the lack of affordable methods for mapping relevant soil attributes is a fundamental problem. The project, of which results are presented in this paper, tries to contribute a module to solve this problem at least to some extent. The project is part of "I4S - Integrated System for Site-Specific Soil Fertility Management" which combines new sensing technologies with dynamic soil-crop models and decision support systems.

The aim of the current investigation is to qualitatively predict the plant available Phosphorus content (Pa) in soil using methods of machine learning based on laboratory data, single sensor data, and sensor data fusion. Machine learning is one of the methods of a so-called black box approach. From the numerous machine learning methods an artificial neural networks (ANN) algorithm was used to carry out the current exploration. The ANN algorithm “learns” relationships between the target and explanatory variables in a model. Several previous works demonstrated that the acidity of soil (pH) enormously influences the phosphorus availability for the plant and improves its prediction. Therefore pH values as an input variable were introduced to improve the predictability of each model. This investigation involves sensors for the apparent electrical resistivity (ERa), the gamma intensity (ɣ) as well as an x-ray fluorescence analyzer (XRF). The results show the following: 1. The performance of qualitative prediction increases from gamma to ERa to XRF; 2. The sensor data fusion approach improves a prediction partly enormously, however not every combination of sensors improves the prediction; 3. The outlook is to examine an estimation based on the results of single sensor data (prediction based on the estimations of single sensor data).