Linking Common Multispectral Vegetation Indices to Hyperspectral Mixture Models: Results from 5 nm, 3 m Airborne Imaging Spectroscopy in a Diverse Agricultural Landscape

TL;DR

This study compares common multispectral vegetation indices with hyperspectral mixture models using airborne data from California agriculture, revealing which indices best estimate green vegetation fractions and their limitations.

Contribution

It provides a quantitative analysis of how well six popular multispectral vegetation indices relate to hyperspectral endmember fractions in diverse agricultural landscapes.

Findings

NIRv, DVI, EVI, EVI2 strongly correlate with green vegetation fraction

EVI and EVI2 closely follow a 1:1 relation with hyperspectral estimates

NDVI and SR show weaker, nonlinear relationships and are sensitive to background and saturation

Abstract

For decades, agronomists have used remote sensing to monitor key crop parameters like biomass, fractional cover, and plant health. Vegetation indices (VIs) are popular for this purpose, primarily leveraging the spectral red edge in multispectral imagery. In contrast, spectral mixture models use the full reflectance spectrum to simultaneously estimate area fractions of multiple endmember materials present within a mixed pixel. Here, we characterize the relationships between hyperspectral endmember fractions and 6 common multispectral VIs in crops & soils of California agriculture. Fractional area of green vegetation (Fv) was estimated directly from 64,000,000 5 nm, 3 to 5 m reflectance spectra compiled from a mosaic of 15 AVIRIS-ng flightlines. Simulated Planet SuperDove reflectance spectra were then derived from the AVIRIS-ng, and used to compute 6 popular VIs (NDVI, NIRv, EVI, EVI2, SR, DVI). Multispectral VIs were compared to hyperspectral Fv using parametric (Pearson correlation, r) and nonparametric (Mutual Information, MI) similarity metrics. 4 VIs (NIRv, DVI, EVI, EVI2) showed strong linear relationships to Fv (r > 0.94; MI > 1.2). NIRv & DVI showed strong interrelation (r > 0.99, MI > 2.4), but deviated significantly from 1:1 relative to Fv. EVI & EVI2 were also strongly interrelated (r > 0.99, MI > 2.3) and more closely followed a 1:1 relation with Fv. In contrast, NDVI & SR showed weaker, nonlinear, heteroskedastic relation to Fv (r < 0.84, MI = 0.69). NDVI showed especially severe sensitivity to substrate background reflectance (-0.05 < NDVI < +0.6 for unvegetated spectra) and saturation (0.2 < Fv < 0.8 for NDVI = 0.7). These direct observational constraints on multispectral VI and hyperspectral mixture model comparability can serve as a quantitative benchmark for agronomic applications in the coming era of increasing spatial & spectral resolution Earth observation.