Vegetation Index Sensitivities and Structural Biases
Protecting and monitoring the state of vegetation over a variety of scales generally incorporates the use of automated methods of processing multispectral satellite remote sensing data to inform decision-making. For instance, observation of boreal forest post-disturbance recovery has become a major research theme in Canada. Of specific interest, Ontario’s managed boreal forest, a resource for sustainable harvest, has been impacted by climate change in recent decades, resulting in alterations and an increase in the frequency and intensity of wildfires.
Vegetation spectral enhancement indices work by mathematically magnifying characteristic spectral properties of target features, often by taking the ratio or difference of contrasting spectral responses from key wavelength bands within the electromagnetic spectrum. We clarify the sensitivity of four vegetation indices with potential usefulness for assessing the status of post-disturbance boreal vegetation by assessing their structural biases: the normalized difference vegetation index (NDVI), normalized burn ratio (NBR), near-infrared vegetation index (VINIR), and the infrared vegetation index (VIIR).
Vegetation index structure is partitioned into a calculation space to model every possible computational output. Spectral samples of the boreal forest were derived from Landsat satellite constellation data using GoogleEarth engine best-available pixel (BAP) compositing in cross-reference with the 2020 Land Cover of Canada. Simulated outputs are visualized according to computational domains for undisturbed and disturbed boreal vegetation types by producing histograms and three-dimensional surface plots characterizing input variables and the output metric surface. Cross-sectional transects (values grouped by difference components that define a linear increase in values) of these surfaces illustrate their local variability with boxplots. A principal component analysis including each vegetation index was performed to summarize variability within each index. The equation structures of VIIR and VINIR result in a steep, concave upward slope towards higher index values in contrast to NDVI and NBR which display more gradual, convex slopes. For VIIR and VINIR, the index value increases exponentially for unit decreases in green or shortwave-infrared while NDVI and NBR increased exponentially for unit increases in near-infrared. Index transect comparisons show that NDVI and NBR have consistent variability across all possible inputs. VINIR and VIIR have increasing variability given higher inputs in red and near-infrared or shortwave- and thermal-infrared respectively. VIIR displayed the largest relative dynamic range, while VINIR displayed the smallest over the input range for boreal vegetation. It is concluded that the mathematical structure of VINIR and VINIR possess substantially different computational domains than NDVI and NBR and exhibit a greater output range, indicating increased sensitivity to vegetation status.