Using a global VNIR soil-spectral library for local soil characterization and landscape modeling in a 2nd-order Uganda watershed

Combining global soil-spectral libraries with local calibration samples has the potential to provide improved visible and near-infrared (VNIR, 400-2500 nm) diffuse reflectance spectroscopy (DRS) soil characterization predictions than with either global or local calibrations alone. In this study, a geographically diverse global soil-spectral library with 4184 samples was augmented with up to 418 local calibration soil samples distributed across a 2nd-order Ugandan watershed to predict the amount of clay-size material (CLAY), soil organic carbon (SOC) and proportion of expansible 2:1 clays (termed montmorillonite or MT in the global library). Stochastic gradient boosted regression trees (BRT) were employed for model construction, with a variety of calibration and validation schemes tested. Using the global library combined with 13- and 14-fold cross-validation by local profile for CLAY and SOC, respectively, yielded dambo/upland RMSD values of 89/68 g kg(-1) for CLAY (N=429/410) and 4.2/2.6 g kg(-1) for SOC (N=272/105). These results were obtained despite the challenge of combining spectral libraries constructed using different spectroradiometers and laboratory reference measurements (total combustion vs. Walkley-Black, hydrometer vs. pipette). Using only the global library, a VNIR-derived index of NIT content was significantly correlated with the square root of X-ray diffraction (XRD) NIT peak intensity for local dambo soils (r(2)=0.52, N=59, p<0.0001), an acceptable result given the semi-quantitative nature of the reference XRD method. Though VNIR predictions did not approach laboratory precision, for soil-landscape modeling VNIR characterization worked remarkably well for clay mineralogy, was adequate for mapping dambo depth to 35% clay, and was insufficiently accurate for SOC mapping. (C) 2007 Published by Elsevier B.V.