Proximal soil sensing is a promising tool to quantify soil properties through spectral data and machine learning algorithms for application in pedological investigations. However, the relationship between pedological characteristics and sensor signals is difficult to be elucidated with machine learning models. This study aimed to establish direct linkages between pedological features and data from proximal sensors, namely portable X-ray fluorescence spectrometry (pXRF) and visible and near-infrared spectroscopy (Vis-NIR). Ten podzolic soil profiles, including seven Spodosols, two Ultisols, and one Inceptisol, were described and sampled from Central Taiwan. Principal component analysis (PCA) and k-means clustering method revealed that the major variances of the pXRF were affected by the amounts of spodic materials and organic carbon derived from plant debris. For Vis-NIR, a remarkable absorption of Fe-oxides was observed in the B horizons, and the major variations of Vis-NIR spectra extracted by PCA represented the influence of metal-complexed organic carbon and clay. The principal components (PC) of pXRF and Vis-NIR can be used to separate soil horizons, especially the E horizons were clearly separated as indicated by its high user’s and producer’s accuracies. However, the B and C horizons were poorly separated by the PCs and k-means clustering method. Moreover, the PCs can be used directly to predict the nanocrystalline Fe and Al and the optical density of the oxalate extract. Ten soil profiles were classified based on the predicted values, and nine out of ten soil profiles were in agreement with their classification using chemical extractions. The proposed pXRF and Vis-NIR approaches predicted the chemical criteria of diagnostic horizons and facilitated fast determination of Spodosols.
Read full abstract