This doctoral dissertation explores the use of hyperspectral remote sensing for predicting the content of heavy metals (Zn, Pb, Cd) in the upper soil layers. The research was conducted in the area northeast of Celje, where prolonged industrial activity has caused significant pollution. We developed a methodology based on linking airborne hyperspectral imagery with laboratory spectroradiometric measurements, as well as chemical and pedological analyses of soil samples. Attention is given to comprehensive data processing – from atmospheric correction, input data analysis, spectral transformations, and dataset preparation for modelling, to the selection of relevant variables. The influence of individual input variables and different preprocessing procedures on the quality of predictive models was assessed using conditional permutation importance. The final prediction of selected heavy metal contents is based on a novel approach using a machine learning method, where additional chemical and pedological variables are used only during the model training phase, while the final predictions rely exclusively on spectral data. The presented methodology constitutes a methodological contribution to the development of hyperspectral approaches for soil pollution monitoring and demonstrates potential for more efficient and cost-effective mapping over larger areas.
|
