Making lasagne with spaghetti: geometric and radiometric corrections for hyperspectral data acquired obliquely from UAVs

Cliffs present some of the most spectacular geological exposures, and can provide detailed and spatially continuous geological data for research and industry applications. However, challenging access has until recently limited our ability to get full value from these valuable outcrops. The surge in uncrewed aerial vehicle (UAV) technology has palliated some of these limitations, allowing for rapid and unprecedentedly detailed (sub-cm resolution) surveying with visible-near (VNIR) and shortwave (SWIR) infrared hyperspectral sensors. UAV-based SWIR-sensors typically use a pushbroom acquisition mode that results in significant distortions due to UAV movement. These must be corrected to derive geometrically accurate results. In this contribution we present an open-source workflow for (1) the geometric correction and back-projection of pushbroom hyperspectral data to derive dense 3-D hyperclouds; (2) removal of illumination effects to derive estimates of reflectance spectra and (3) the application of various hyperspectral mapping techniques to extract lithological and mineralogical information. This workflow is implemented in the open-source python package hylite to facilitate and encourage future research and open access science by researchers and industry.
Our approach is different to the correction workflows implemented by camera vendors (e.g., PARGE) as it directly associates points in a photogrammetric point cloud with pixels in the hyperspectral pushbroom image. The resulting mapping matrix captures the “many to many” relationship between points and pixels. For pushbroom imagery a single point can be visible from several pixels, and each pixel will contain multiple points, and that facilitates the transfer and fusion of hyperspectral data onto the geometrically accurate point cloud. This true-3D approach is essential in areas of complex relief, such as cliffs or open-pit mines, as these geometries cannot be projected onto a 2-D image plane (orthomosaic) without significant distortion and geometric errors.
Additional advantages of this approach are: (1) high resolution panchromatic data from the photogrammetric point cloud can be used to automatically correct for sensor boresight, and (2) the topographic information captured by the point cloud provides the geometric information (e.g., surface orientation and skyview factor) required to correct for illumination effects and derive reflectance spectra. The resulting reflectance hypercloud can then be analysed using a variety of methods implemented in hylite (e.g., minimum wavelength mapping, band ratio calculation or spectral unmixing) to create objective and reproducible maps of lithology or mineralogy.