Estimation of soil organic carbon using hyperspectral remote sensing data and a large scale soil spectral library - from laboratory to space

Our soils are the largest terrestrial carbon storage of the planet. With proceeding climate change and to retain food security for a growing human population, the pressure rises to increase the knowledge about the soil’s current status and especially its carbon content. Mapping the status of the soil carbon content presents a challenge, especially for larger areas, since conventional methods require extensive sample collection and expensive laboratory analyses. To fully utilise the soil’s capacity and preserve the its health in the future, it is required to monitor key soil parameters thereby adding a temporal dimension. Therefore, suitable approaches are needed to estimate the carbon content regularly and on larger scales. Soil spectroscopy has the potential to support this aim. At laboratory and airborne scale soil spectroscopy already proved to be an accurate method to estimate certain soil properties. Recently, advanced optical hyperspectral spaceborne imaging sensors have become available that have the potential to create quantitative soil carbon maps. These Earth observation (EO) sensors can also be a future monitoring tool to e.g. combat soil degradation. In this thesis, approaches were developed and applied to evaluate these new hyperspectral EO sensors to map the spectrally active soil organic carbon (SOC) content at different scales. Ground reference data represent a bottle neck but are essential for model generation and evaluation. A potential solution is the additional usage of existing large scale soil spectral libraries (SSLs) which contain information on physiochemical and spectral properties. Thereby, the challenge is the decreasing model accuracy with increasing size of the study site. Therefore, the first aim was to improve the prediction accuracy of SOC content using laboratory spectra and the European wide LUCAS SSL. Especially a memory-based learning approach (local partial least square regression: local PLSR) could improve the prediction accura