Hyperspectral Imaging for the Quantification of Soil Organic Carbon—a Proximal Sensing Experiment Reflecting the Importance of Using Pure Soil Data in Vis-NIR-SWIR Soil Spectroscopy

Abstract Proximally sensed hyperspectral images of soil surfaces provide fine-scale pixel information that might be used for the removal of spectrally disturbing effects induced by roughness or the presence of non-soil materials and thus could allow a better quantification of soil variables. For a set of 50 soil surfaces from which we took samples in an undisturbed condition, we tested the usability of HySpex VNIR-1800 – SWIR-384 scans performed in the lab to estimate soil organic carbon (SOC). Strategies to compensate disturbances were (i) the application of different spectra pre-processing techniques (derivatives, normalization, orthogonal signal correction), and (ii) the use of sub-image information defined by regular gridding or spectral unmixing (SU), the latter to remove non-soil pixels. Unprocessed image data (mean absorbances) allowed only a poor SOC estimation in a 10-fold cross-validation (RMSE = 5.03 g kg −1 , R 2 = 0.36, RPD = 1.26, RPIQ = 1.21). Marked improvements were obtained with the use of pure soil pixels, identified with SU based on an unsupervised endmember definition. The additional usage of an ensemble of different pre-processing methods further improved results slightly to an RMSE that finally equalled 3.68 g kg −1 and an R 2 at 0.66 (RPD = 1.73, RPIQ = 1.65). Our results underline the importance of using soil data that is as pure as possible for the spectral retrieval of key soil variables from, in our case, proximally sensed image data. The approach can be in principle transferred to remote sensing data, given that enough bare soil pixels can be identified for the plot, field or region that is studied.