Using legacy data for correction of soil surface clay content predicted from VNIR/SWIR hyperspectral airborne images

Visible, near-infrared and short-wave infrared (VNIR/SWIR, 0.4-2.5 mu m) hyperspectral airborne imaging has been demonstrated to be a potential tool for topsoil property mapping (such as free iron, clay, and organic matter) over bare soils of large areas. Nevertheless, one of the limiting factors of hyperspectral airborne data use for soil property mapping is the need for a set of soil spectra extracted from bare soils pixels of the VNIR/SWIR airborne data and the corresponding soil property values measured over soil samples collected over the bare soils pixels for which soil spectra are extracted. We propose to test a new approach which uses legacy soil data collected over and/or around the study site, instead of soil property values measured over soil samples collected over bare soils pixels. As legacy soil samples can be inaccurately localized or can be located out of bare soils of hyperspectral airborne data or out of the study area, these data could be unusable as calibration data for classical predictive models (such as the partial least-squares regression method). So the proposed approach first uses a spectral clay index to estimate clay contents (in relative values as it is done without calibration) and then transform these estimated clay contents thanks to a correction of the distribution and range of clay content estimations using legacy soil data. This procedure is compared to a linear model built from the spectral clay index and calibrated using a reference database. The spectral index was proposed by Levin et al. (2007) using spectral bands at 2205, 2.13, 2.224 mu m. This study employs the VNIR/SWIR AISA-DUAL hyperspectral airborne data acquired over an area of 300 km(2) in a Mediterranean region. Our results show that 1) this spectral index offers predictions with low accuracy in terms of the coefficient of determination, R-2, which is associated with high bias and SEP; 2) the distribution and range correction made using legacy soil data allows for both an increase of accuracy (R-2) and an improvement of bias and SEP; 3) it is better to have a small number of legacy ground measurements focused on the study area than a high number of legacy ground measurements dispersed on and far from the study area; 4) the correction of the prediction bias is highly dependent on the legacy soil data quality; and 5) regardless of which legacy soil database is used, the soil pattern is discriminated. With the coming availability of the next generation of hyperspectral VNIR/SWIR satellite data for the entire globe, this study may open a new way toward accessible and cheap methods for the delivery of soil property maps to the geoscience community. (C) 2016 Elsevier B.V. All rights reserved.