Soil spectroscopy represents a low-cost alternative to routine time-consuming and expensive laboratory analyses. Its ability to measure a wide range of different chemical and physical soil properties was shown previously in many studies. Particularly, for organic carbon content, a reliable prediction accuracy is usually achieved. This is due to strong spectral signature of soil organic carbon and other distinct spectral implications of soil characteristics strongly tied to it, e.g. soil colour. All the known studies, however, deal with situation where the study area is fully covered (either in the manner of design- or model-based sampling approach) with calibration points. But in many cases the sampling strategy was initially designed for other purposes, falling outside requirements of spectroscopy for proper model calibration. Hence, here we attempt to test the ability of soil spectroscopy in the situation when only a minor isolated part (the steepest one) of the study area was sampled for calibration points, and predictions were made for its several time larger surroundings. For model training we used Partial Least Squares Regression (PLSR) technique and four different spectra pre-treatment methods (Savitzky-Golay smoothing, first and second derivative, and baseline normalization via continuum removal). Results show high potential (R2 ≈ 0.70-0.80) of the method for rough terrain landscapes strongly affected by water erosion, even if the distance from calibration to prediction points is large.
Partial least Squares Regression, Spectra Pre-Treatment, Soil Assessment.