A Spectral Assignment-Oriented Approach to Improve Interpretability and Accuracy of Proxy Spectral-Based Models

Abstract

In modeling chemical attributes using hyperspectral data, nonlinear relationships between the predictor and the response are frequent. The common nonlinear modeling techniques improve prediction accuracy but suffer from low interpretability of the models. In this paper, we demonstrate a new multivariate modeling method, denoted as spectral assignment-oriented partial least squares (SAO-PLS), which is designed to provide a nonlinear modeling solution with strong interpretability products. The need for this approach is apparent when a given sample population consists of different spectral features for different levels of the response. Accordingly, the suggested SAO-PLS algorithm segments the data in an optimal location on the response distribution by maximizing the difference in spectral assignments between two clusters. SAO-PLS is applied here to two test cases with different characteristics: 1) an established data set containing airborne hyperspectral data of asphaltic roads, merged with in situ measured dynamic friction values captured using a standardized method and 2) a soil spectral library, spectrally measured with an analytical spectral device spectrometer, to which organic carbon measurements were applied. Our results demonstrate the superiority of SAO-PLS over partial least-squares regression for both model accuracy and interpretability, providing a deeper understanding of the underlying processes.