Unsupervised clustering of soil spectral curves to obtain their stronger correlation with soil properties

Abstract

The laboratory measurements of the diffuse spectral reflectance of 212 soil samples, representing many various taxonomic units, collected throughout the area of arable lands in Poland, were conducted to investigate the relationship between the soil reflectance and their selected properties. It was found that among various tested transformations the first derivative of the soil reflectance was the one most strongly correlated with the content of textural fractions, soil organic carbon, Fe and CaCO3, The use of unsupervised Ward's Euclidian distance based on clustering algorithm to split the total dataset into subsets, according to the shape and the level of the soil spectra, improved the correlation between soil properties and the transformed spectral data. The highest values of the coefficient of determination R2 for clay and Fe contents on the total dataset reached only 0.64 and 0.56, respectively. Using the ED Ward's algorithm, six subsets were formed and their R2 increased up to 0.87 and 0.80, respectively.