Illumination correction for close-range hyperspectral images using spectral invariants and random forest regression

Abstract

Identifying materials and retrieving their properties from spectral imagery is based on their spectral reflectance calculated from the ratio of reflected radiance to the incident irradiance. However, obtaining the true reflectances of materials within a vegetation canopy is challenging given the varying illumination conditions across the canopy – i.e., the irradiance incident on a surface inside the canopy – caused by its complex 3D structure. Instead, in remote sensing, reflectances are calculated from the ratio of the spectral radiance measured by the sensor to the top-of-canopy (TOC) spectral irradiance, resulting in apparent reflectances that can significantly differ from the true reflectance spectra. To address this issue, we present a physically based illumination correction method for retrieving the true reflectances from close-range hyperspectral TOC reflectance images. The method uses five spectral invariant parameters to predict the illumination conditions from TOC reflectance and compute the corrected spectrum using a physically based model. For computational efficiency, the spectrally invariant parameters were retrieved using random forest regression trained with Monte Carlo ray tracing simulations. The method was tested on close-range imaging spectroscopy data from dense and sparse vegetation canopies for which reference in situ spectral measurements were available. This work is a step toward resolving the 3D radiation regime in vegetation canopies from TOC hyperspectral imagery. The retrieved spectral invariants provide a physical connection to the structure of the observed vegetation canopy. The true spectra of artificial and natural materials in a vegetation canopy, determined under various illumination conditions, allow their more robust (bio)chemical characterization, opening new applications in vegetation monitoring and material detection, and machine learning makes it possible to apply the method rapidly to large hyperspectral image sets.