rs-local data-mines information from spectral libraries to improve local calibrations

Abstract

Diffuse reflectance spectroscopy in the visible–near infrared (vis–NIR) and mid infrared (mid-IR) can be used to estimate soil properties, such as organic carbon (C) content. Compared with conventional laboratory methods, it enables practical and inexpensive measurements at finer spatial and temporal resolutions, which are needed to improve the assessment and management of soil and the environment. Measurements of soil properties with spectra require empirical calibration and soil spectral libraries (SSL) have been developed for this purpose at the regional, continental and global scales. Calibrations derived with these SSLs, however, are often shown to predict poorly at local sites. Here we present a new method, rs-local, that uses a small representative set of site-specific (or ‘local’) data and re-sampling techniques to select a subset of data from a large vis-NIR SSL to improve calibrations at the site. We demonstrate the implementation of rs-local by estimating soil organic C in two fields with different soil types, one in Australia and one in New Zealand. We found that with as few as 12 to 20 site-specific samples and the SSL, training datasets derived with rs-local could accurately predict soil organic C concentrations. Predictions with the rs-local data were comparable to, or better than those made with site-specific calibrations with up to 300 samples. Our method outperformed other published ‘local’ spectroscopic techniques that we tested. Thus, rs-local can effectively improve both the accuracy and financial viability of soil spectroscopy. Highlights: We describe a new algorithm (rs-local) for site-specific calibration using existing spectral libraries. rs-local is a data driven method that makes no assumptions on spectral or sample similarities. rs-local improved the accuracy of soil organic carbon estimates using spectroscopy. rs-local improves the economic viability of soil spectroscopy.