Snapshot Hyperspectral Imaging for Soil Diagnostics -- Results of a Case Study in the Spectral Laboratory

Abstract

Field reflectance spectroscopy has been widely used in proximal soil sensing. Results of spectroscopic approaches depend, inter alia, from the experimental setup and the applied spectroradiometric instrumentation. Beyond the traditional instrument concepts (acquisition of ground truth data with field spectroradiometers, air- and spaceborne scanners), there are currently alternative developments in the ground-based or near-ground spectroscopy: The hand-held and thus mobile non-scanning hyperspectral imaging technique might be one previously missing part in the operational spectral data chain to be used for down- and upscaling purposes. It should effectively bridge the gap between point and image data as it enables a very rapid data acquisition. This study describes how readings of a hyperspectral frame camera (in the nominal spectral range from 450 nm to 950 am) could be utilised for soil detection and analysis. The proximally sensed hyperspectral images were compared to 1D spectroradiometric data, both acquired in the lab using raw, sieved and grinded soil samples. Measured spectral datasets were then used to define multivariate calibration models, i.e., the spectra were analysed to extract quantitative models between spectral data and soil constituents of interest determined by wet chemical analysis. We used partial least squares regression (PLS) as statistical calibration method to estimate soil organic carbon (OC), hot-water extractable carbon (HWE-C) and nitrogen (N). The results that we obtained from the camera data were satisfactory (with coefficients of determination (R-2) between 0.62 and 0.84 in the cross-validation), but only with crushed samples and when combining PLS with CARS (competitive adaptive reweighted sampling), an effective spectral variable selection technique. For in-field studies without any sample preparation, stratified approaches considering soil surface roughness and/or the elimination of shadow pixels from the acquired images might both be promising to improve the accuracy of obtainable estimates.