PE&RS August 2014 - page 706

Hyperspectral Narrowband
Combinations
Various 4, 6, 8, 10, 12, 16, and 20 best HNB combinations (Table
4) can be required to compare with various corresponding
broadband data available to us such as the 9 non-thermal
bands of Landsat-8 OLI, 4 band IKONOS, and 4 band IRS.
Meta-analysis of literature (e.g., Thenkabail et al., 2011b;
Thenkabail et al., 2013) indicates various HNB best-band
combinations (Table 4). These HNB combinations work the
best in classifying or modeling vegetation or agricultural crops
when they come from various portions of the spectrum (e.g.,
visible, near infrared, shortwave infrared). The number of
HNB bands to use and their combinations will depend on the
complexity of vegetation or crop types involved. For example,
in order to classify 2 crop types over a small area with high
degree of accuracy, only the best 4 bands may suffice. But
when multiple crops are involved, 16 or 20 bands or even all
28 bands (Table 2, Figure 4a) maybe required. In modeling
biophysical or biochemical quantities of vegetation or crops, one
can compose HVIs (Table 3) based on two band combinations
(Table 3) or multiple bands (Table 2, Table 4). It is possible to
establish multiple HNB band based indices in modeling crop
or vegetation biophysical or biochemical quantities. However,
past researches (Thenkabail et al., 2000; Thenkabail et al.,
2002; Thenkabail et al., 2004a; Thenkabail et al., 2004b; Bian
et al., 2010; Clark and Roberts, 2012; Mariotto et al., 2013;
Marshall and Thenkabail, 2014) have shown that R-square
values are maximum anywhere between the use of ~3 to 8
bands (Thenkabail et al., 2004b; Mariotto et al., 2013; Marshall
and Thenkabail, 2014), beyond which the relationship between
the number of HNBs and R-square is asymptotic.
Conclusions
A summary of the strengths, limitations, and challenges
involved in hyperspectral remote sensing (or imaging
spectroscopy) of vegetation and agricultural crops is provided
in this paper.
The paper identifies optimal HNB- centers and widths
(Figure 4b, Table 2) and HVIs (Table 4) that are best for
classifying, separating, monitoring, modeling, and mapping
vegetation and agricultural crops. Overall, ~15 to 20, but
no more than about 28 HNBs (Table 2) provide optimal
information in vegetation or crop classification. Typically,
HNBs achieve about ~30% higher accuracies compared to 6
non-thermal broadbands in classifying 5 to 12 vegetation or
crop categories. Beyond these optimal HNBs, the accuracies
asymptote with an increase in the number of HNBs.
There are specific HVIs (Table 3) that best characterize
and model vegetation and crop biophysical and biochemical
properties. These HVIs are grouped into 6 distinct categories:
1. Hyperspectral biomass and structural indices (HBSIs),
2. Hyperspectral biochemical indices (HBCIs),
3. Hyperspectral red-edge indices (HREIs),
4. Hyperspectral water and moisture indices (HWBIs),
5. Hyperspectral light-use efficiency index (HLUEI), and
6. Hyperspectral lignin-cellulose index (HLCI).
It must be noted, that the use of the first index from each of
the six categories is, typically, the best index for the category
Table 4. Best hyperspectral narrowband (HNB) combinations.
Best 4 bands 550, 682, 855, 970
Best 6 bands 550, 682, 855, 970, 1075, 1450
Best 8 bands 550, 682, 855, 970, 1075, 1180, 1450, 2205
Best 10 bands 550, 682, 720, 855, 970, 1075, 1180, 1245, 1450, 2205
Best 12 bands 550, 682, 720, 855, 910, 970, 1075, 1180, 1245, 1450, 1650, 2205
Best 16 bands 490, 515, 550, 570, 682, 720, 855, 910, 970, 1075, 1180, 1245, 1450, 1650, 1950, 2205
Best 20 bands 490, 515, 531, 550, 570, 682, 720, 855, 910, 970, 1075, 1180, 1245, 1450, 1650, 1725, 1950, 2205, 2260, 2359
“It is obvious that there is a need for
Whole Spectral Analysis (WSA) as well
as Optimal Spectral Analysis (OSA)”
706 August 2014 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING