Assessment of soil water retention characteristics based on VNIR/SWIR hyperspectral imaging of soil surface

Abstract

This paper aims to create supervised classification models of the soil water potential based on hyperspectral data of Polish mineral soils (104 samples) from the Visible and Near-InfraRed (VNIR) and Short-Wave InfraRed (SWIR) range and selected soil physico-chemical properties, such as organic carbon content, or fraction of sand, silt, and clay. Soil water content regression models were also created, which took into account the soil water potential ranging from 98.1 J∙m-3 to 1554249 J∙m-3. Several machine learning algorithms were tested to create models of the soil water potential and the soil water content. It occurred that reflectance characteristics of the soils exhibit a high correlation with soil moisture. Gaussian Processes (GP) model was most suitable for the estimation of the soil water content, regardless if input data contained pure reflectance spectra (R=0.82), or if they were supplemented with selected physico-chemical soil properties and soil water potential (R=0.94). No improvement of the models’ accuracies was noticed when only the selected soil physical and chemical properties were included as inputs, which suggested that the soil surface spectral data contained in themselves information, which strictly belonged to specific soil physico-chemical properties. Among classification models of the soil water potential, the LOG method had the highest percentage of correctly classified cases. More than 65% of all cases were correctly classified if the spectral data, moisture, and other properties of the tested material were included, and more than 54% when the independent variables did not include soil moisture. The majority of misclassified cases were by one class. The greatest accuracy was achieved for the two extreme values of pF (0 and 4.2), while the worst one was for pF2.2 and pF2.7.