Abstract
Soils are the largest terrestrial carbon pool providing important ecosystem services, essential for sustainable land management, agricultural productivity, and climate regulation. Spaceborne imaging spectrometers represent a promising tool for producing up-to-date, inexpensive and spatially explicit maps. The recent availability of spaceborne imaging spectroscopy sensors provide high signal-to-noise ratio, such as the EnMAP, offers new opportunities for detailed and spatially explicit soil assessment.
This study investigates the potential of EnMAP hyperspectral imagery for retrieving and mapping key topsoil properties, with a particular focus on soil organic carbon (SOC), using advanced machine learning and spectral processing techniques. The analysis conducted across agricultural landscapes in Kopaida region in Central Greece and was supported by an extensive soil sampling campaign designed to capture the spatial variability of soil conditions within the study area.
Bare soil spectra were extracted from EnMAP imagery, and the effects of spectral transformations and feature selection on model performance were assessed. Random Forest (RF) models were trained based on different spectral transformations and feature-selection methods, and their impacts on predictive performance were compared to derive a robust modelling workflow. The SOC results showed strong performance at the laboratory scale. Results demonstrated the potential of EnMAP hyperspectral observations for operational soil monitoring and support the development of scalable methodologies for regional soil monitoring and digital soil mapping. Overall, the generated soil maps revealed substantial spatial variability across the study area, highlighting the influence of local environmental factors on soil distribution patterns.