Feature extraction for hyperspectral mineral domain mapping: A test of conventional and innovative methods

Hyperspectral (HS) imaging holds great potential for the mapping of geological targets. Innovative acquisition modes such as drone-borne or terrestrial remote sensing open up new scales and angles of observation, which allow to analyze small-scale, vertical, or difficult-to-access outcrops. A variety of available sensors operating in different spectral ranges can provide information about the abundance and spatial location of various geologic materials. However geological outcrops are inherently uneven and spectrally heterogeneous, may be covered by dust, lichen or weathering crusts, or contain spectrally indistinct objects, which is why classifications or domain mapping approaches are often used in geoscientific and mineral exploration applications as a means to discriminate mineral associations (e.g. ore or alteration zones) based on overall variations in HS data. Feature extraction (FE) algorithms are prominently used as a preparatory step to identify the first order variations within the data and, simultaneously, reduce noise and data dimensionality. The most established FE algorithms in geosciences are, by far, Principal Component Analysis (PCA) and Minimum Noise Fraction (MNF). Major progress has been conducted in the image processing community within the last decades, yielding innovative FE methods that incorporate spatial information for smoother and more accurate classification results. In this paper, we test the applicability of conventional (PCA, MNF) and innovative FE techniques (OTVCA: Orthogonal total variation component analysis and WSRRR: Wavelet-based sparse reduced-rank regression) on three case studies from geological HS mapping campaigns, including drone-borne mineral exploration, terrestrial paleoseismic outcrop scanning and thermal HS lithological mapping. This allows us to explore the performance of different FE approaches on complex geological data with sparse or partly inaccurate validation data. For all case studies, we demonstrate advantages of innovative FE algorithms in terms of classification accuracy and geological interpretability. We promote the use of advanced image processing methods for applications in geoscience and mineral exploration as a tool to support geological mapping activities.