Analysis of the uncertainties affecting predictions of clay contents from VNIR/SWIR hyperspectral data

Visible, Near-Infrared and Short Wave Infrared (VNIR/SWIR, 350-2500 nm) hyperspectral imaging spectroscopy can be used to develop maps of estimated soil properties. The performance of the estimations obtained from regression models is usually assessed with figures of merit such as the standard error of calibration, the standard error of prediction or the ratio of performance deviation. All of these parameters are estimated during the model building and validation stages to evaluate the global model performance. Beyond these global indicators, the evaluation of the uncertainty that affects predictions is a major trend in analytical chemistry and chemometrics but not yet in hyperspectral imagery. Several approximate expressions and resampling methods have been proposed to estimate the prediction uncertainty when using multivariate calibrations from laboratory spectra. Based on these propositions, this paper studies a process of mapping and analyzing the uncertainties that affect soil property predictions obtained from VNIR/SWIR airborne data using several methods. An application to real VNIR/SWIR airborne data of clay content was used to compare the methods. The different cases yielded insights into the sources of uncertainty and showed that uncertainty analysis can guide the user to better sampling, better calibration and ultimately better mapping.