GLOBAL-LOCAL: A new approach for local predictions of soil organic carbon content using large soil spectral libraries

There is much interest in the use of large visible near infrared (vis-NIR) soil spectral libraries for rapid soil analysis and to further soils research. However, the use of such large and diverse spectral libraries often provides biased predictions at local scales. Here, we present a new approach, GLOBAL-LOCAL, which aims to use large libraries to develop spectroscopic models that can fit locally. This is achieved by selecting the most common spectral neighbours of a small representative set (Lab) from the target site using a combination of principal component analysis (PCA) and k-nearest neighbours, followed by selection of the calibration with the lowest root mean square error (RMSE) in predicting the Lab, which is then added to the calibration to derive the GLOBAL-LOCAL. We evaluated this approach by predicting the soil organic carbon (SOC) content in two different target sites, one in Canada (L & eacute;vis, QC; N = 111) and the other in Finland (Maaninka; N = 101). Partial least squares regression (PLSR) was used for model development and validation. The soil spectral library used was a combination of that developed by the World Agroforestry Center and International Soil Reference Information Centre (N = 3875), and that developed at the Quebec Research and Development Centre (N = 1051). Based on a preliminary analysis, we selected 25 Lab samples from each target site using conditioned Latin hypercube sampling (cLHS). Prediction accuracy by GLOBAL-LOCAL for L & eacute;vis and Maaninka (RMSE = 1.41 and 3.08 g kg(-1), RPIQ = 5.9 and 4.1, respectively) were better than using the global calibration (RMSE = 20.3 and 23.6 g kg(-1), RPIQ = 0.41 and 0.53, respectively). The GLOBAL-LOCAL was better or similar to calibrations with site-specific samples (RMSE = 2.38 and 4.68 g kg(-1), RPIQ = 3.5 and 2.7, respectively) and Lab samples only (RMSE = 1.41 and 2.54 g kg(-1), RPIQ = 5.9 and 4.9, for L & eacute;vis and Maaninka, respectively). In addition, GLOBAL-LOCAL outperformed three other approaches, RS-LOCAL, the spectrum-based learner and LOCAL. While 25 Lab samples was regarded as optimum by cLHS, further examination revealed that as little as 10 Lab samples may be sufficient for the GLOBAL-LOCAL. The method uses the information in the large spectral library so that it can substantially reduce analytical cost and greatly improve predictions on local scales.