A remote sensing adapted approach for soil organic carbon prediction based on the spectrally clustered LUCAS soil database

The estimation of the soil organic carbon (SOC) content plays an important role for carbon sequestration in the context of climate change, food security and soil degradation. Reflectance spectroscopy has proven to be a promising technique for SOC quantification in the laboratory and increasingly from air- and spaceborne platforms, where hyperspectral imagery provides great potential for mapping SOC on larger scales with regular updates. When applied on larger scales, soil prediction accuracy decreases due to the inhomogeneity of samples. In this paper, we examined if spectral clustering of the LUCAS EU-wide topsoil database is successful without using other covariates than the spectral database and can improve SOC model performance compared to a reference model that was calibrated on the whole database without clustering. Different clustering methodologies were tested, including a k-means clustering based on principal component analysis or based on spectral feature variables, combined with partial least squares regression (PLSR) models, and a clustering based on a local PLSR approach which builds a different multivariate model for each sample to be predicted. Furthermore, in order to allow for subsequent application to hyperspectral remote sensing data, atmospheric water wavelengths were removed from the analyses. The local PLSR approach achieved best results and was additionally applied to LUCAS spectra resampled to the upcoming hyperspectral EnMAP sensor which led to good results: R-2 = 0.66, RMSEP = 5.78 g kg(-1) and RPIQ = 1.93. The k-means clustering approach showed slightly better results than the reference model. Overall, our results showed similar performances for SOC prediction models compared to other approaches using PLSR with a larger spectral range and other soil parameters as covariates. This study shows that (i) it is possible to transfer the local PLSR approach onto a wavelengths reduced spectral library and to predict estimations of SOC at low-cost with reasonable accuracy based on large scale soil databases; and (ii) that the local regression approach is a valuable tool for SOC prediction models based solely on spectral data without the use of other soil covariates.