Machine learning approaches for soil classification in a multi-agent deficit irrigation control system

We propose a novel approach to automating soil texture classification from in situ sensors in the field. This approach exploits the features of a soil water retention model using machine learning algorithms. Knowledge of the soil textures is then used to learn the composition of the field and its soil horizons. We discuss the role of soil texture classification within our multi-agent irrigation control system and then conduct a preliminary experiment with soil water retention data from the UNSODA database. The system is evaluated with respect to six classifiers. A maximum classification rate of 85.11% was achieved with a MLP neural network, although performance was relatively consistent across all classifiers.