Soil moisture prediction with feature selection using a neural network

For the problem of soil moisture prediction, existing approaches in literature [6, 11] usually utilize as many decision factors as possible, e.g. rainfall, solar irradiance, drainage, etc. However, the redundancy aspect of the decision factors has not been studied rigorously. Previous research work in data mining has shown that removing redundant features improves rather than deteriorates the prediction accuracy. In this paper, we propose an approach to the problem of soil moisture prediction, which integrates two components: feature selection and prediction model: A method is proposed for feature selection that effectively removes the redundant decision factors; This is followed by a feedforward neural network to make prediction based on the retained (i.e. non-redundant) decision factors. Empirical simulations demonstrate the effectiveness of the proposed approach. In particular, with the help of the proposed feature selection component to remove redundant decision factors, the proposed approach is shown to give better prediction accuracy with lower data collection cost.