Dimensionality
Analysis
2.1. Intrinsic Dimensionality and
Relevance to Arctic Optical Diversity and Ecosystems
Intrinsic dimensionality, the number of independent degrees of freedom
in a dataset, has been used to measure the information content of
spectral catalogues (Cawse-Nicholson et al., 2021; Thompson et al.,
2017). Roughly speaking, the dimensionality of the upwelling light field
indicates different physical and chemical properties apparent in the
terrain that are revealed in the measured radiance spectrum. Here we
characterize the differences in intrinsic dimensionality among different
areas of the Arctic, as represented in the airborne ABoVE dataset
acquired by AVIRIS-NG over Alaska and northwestern Canada. This
dimensionality analysis demonstrates the high degree of spectral
diversity of the Arctic tundra, highlighting the advantage of a
large-scale experiment such as ABoVE and the increased information
content provided by imaging spectrometers, as opposed to multispectral
sensors.
2.2. Dimensionality Analysis
Approach
We analyzed the AVIRIS-NG dataset acquired in 2017, consisting of over
200 different flightlines, segmented at ~3 km intervals
(600 pixels at 5 m). The measured spectrum is calibrated to units of
absolute radiance as in (Chapman et al., 2019). We estimated surface
reflectance spectra using the approach of (Thompson et al., 2018).
Finally, we calculated the intrinsic dimensionality of each segment
independently using the strategy of (Thompson et al., 2017). Within each
segment, the intrinsic dimensionality was calculated from the image
stack, cloud fraction and the mean and standard deviation of Normalized
Difference of Vegetation Index (NDVI) were summarized from the imagery,
and the central latitude and longitude were extracted. We plotted the
frequency distribution of dimensionality for the cloud-free segments,
summarized by latitude and NDVI, to examine trends and patterns in
spectral dimensionality (Figure 2).
2.3. Dimensionality Analysis Results
and
Implications
Dimensionality was calculated for a total of 14,519 segments, of which
12,626 were cloud-free and used in subsequent analysis. Dimensionality
values were positively skewed with a long tail of high values.
Generally, a broad range of dimensionality was observed across the
gradient of latitude and greenness. Above 62° N, segments with moderate
NDVI values (0.25-0.75) consistently had higher dimensionality than
those with either low (< 0.25) or high (> 0.75)
NDVI. The lowest dimensionality values, < 20, were found
mostly in the low NDVI category corresponding to non-vegetated terrain
and open water. These systems were optically less diverse than the
vegetated areas. Inconsistent observing conditions, such as solar angle
and the amount of atmospheric haze, affect the sensor’s ability to
resolve the subtlest features and probably play some role in the broad
spread of dimensionality values. Even excluding the largest values, the
modes of the distributions lie between 20 and 40, similar to previous
studies of midlatitude regions (Thompson et al., 2017). This
demonstrates that Arctic tundra exhibits considerable spectral
heterogeneity across the surveyed region. Unlocking the large amount of
information available in these dimensions can provide new insights into
tundra characteristics and function and will be the focus of future
studies. Considering that this analysis was restricted to one segment
size, it is quite likely that there is even more information embedded in
these spectra. Dimensionality analyses like this but conducted across a
range of segment sizes and with coincident finer-grained data provide an
important opportunity to inspect the spatial resolution of vegetation or
surface substrate patches in Arctic landscapes. Such analyses may be
necessary to understand the properties of interest and heterogeneity
within the mixture of non-vegetated and vegetated surfaces.