Comparison of spectral selection methods in the development of classification models from visible near infrared hyperspectral imaging data

Applications of hyperspectral imaging (HSI) to the quantitative and qualitative measurement of samples have grown widely in recent years, due mainly to the improved performance and lower cost of imaging spectroscopy instrumentation. Data sampling is a crucial yet often overlooked step in hyperspectral image analysis, which impacts the subsequent results and their interpretation. In the selection of pixel spectra for the calibration of classification models, the spatial information in HSI data can be exploited. In this paper, a variety of sampling strategies for selection of pixel spectra are presented, exemplified through five case studies. The strategies are compared in terms of the proportion of global variability captured, practicality and predictive model performance. The use of variographic analysis as a guide to the spatial segmentation prior to sampling leads to the selection of representative subsets while reducing the variation in model performance parameters over repeated random selection.