Application of non-linear statistical tools to a novel microwave dipole antenna moisture soil sensor

In this paper we will show the boosting performance of nonlinear machine learning techniques applied to a novel soil moisture sensing approach. A probe consisting in a transmitting and a receiving dipole antenna was set up to indirect assess the moisture content (%) of three different types of soils (silty clay loam, river sand and lightweight expanded clay aggregate, LECA). Gain and phase signals acquired in the 1.0 GHz - 2.7 GHz frequency range were used to built predictive models based on linear PLS regression and on nonlinear Kernel-based orthogonal projections to latent structures (K-OPLS) algorithms. K-OPLS algorithm slightly increased the accuracy of the models built on the gain response on specific kind of soils with respect to classical linear PLS. However, the predictability increases significantly in the case where the models are built starting from a matrix containing all the considered soil samples (silty clay loam + river sand + LECA) achieving R2 = 0.971 (RMSE = 1.4%) when using K-OPLS non-linear model with respect to R2 = 0.513 (RMSE = 6.1%) obtained using linear PLS. Therefore, K-OPLS algorithm appears to give a significant improvement to modelling data where nonlinear behaviours occur.