Asghar Ghorbani

This study presents a novel method for automatic soil texture classification that combines frequency domain dielectric spectroscopy (2–500 MHz) with supervised Machine Learning (ML). Broadband dielectric spectra were measured across five soil textures within a range of controlled water content- and temperature-conditions using a coaxial cavity probe and vector network analyzer. Multiple supervised ML models and spectral feature configurations were evaluated. The ensemble subspace discriminant (ESD) model achieved the highest performance, with up to 90% classification accuracy under random cross-validation. Support vector machines (SVM) and artificial neural networks (ANNs) also performed strongly, particularly with reduced or engineered feature sets. To evaluate model generalization, we implemented a group-aware cross-validation approach that ensured all spectra from a given soil and moisture condition were assigned exclusively to either training or test sets, avoiding overlap of similar samples. This more conservative strategy yielded accuracies up to 77% when dry or water saturated samples were excluded from the test set. Classification performance was highest under moist, unsaturated conditions, where dielectric dispersion most clearly reflected textural differences. These findings demonstrate that ML models can extract soil texture-related patterns from dielectric spectra, offering a viable method for real-time, in situ soil classification with potential benefits for precision irrigation, high-resolution soil mapping, and dielectric soil moisture sensor accuracy.