50% abundance of the respective end-member with less than
20% infeasibility.
Subsequently, individual classification images were
brought together so as to obtain an overall classification image
of the analyzed outcrop. The results of the
MTMF
classifica-
tion showed that combining spectral data from
VNIR
and
SWIR
ground-based hyperspectral cameras provided the opportuni-
ty to identify and differentiate hematitic limestone from lime
mudstone exposed in the analyzed outcrop, which otherwise
was not possible due to their similar spectral response in
the
SWIR
range (Figure 7C and 7E). The classification accura-
cies of
VNIR
+
SWIR
and
SWIR
images were also evaluated and
compared based on overall and producer accuracies generated
from comparison to the manually classified reference images
guided by reflectance spectroscopy results and field observa-
tions (Table 3).
Supporting the qualitative image assessment, inspection of
the classification accuracies indicated that the
VNIR
+
SWIR
im-
age classification is improved over the
SWIR
image classifica-
tion. The overall classification accuracy increased from 68%
to 91% whereas the Kappa coefficient increased from 0.61 to
0.89. While producer accuracies of each class have increased,
the most significant increase in producer accuracy was in
lime mudstone class. Moreover,
VNIR
+
SWIR
image classifica-
tion resulted in markedly fewer number of unclassified pixels
regardless of the class.
Conclusions
This work was an attempt to combine spectral from panoram-
ic
VNIR
(0.4−1.0 µm) and
SWIR
(1.0−2.5 µm) ground-based hy-
perspectral cameras to obtain continuous
VNIR
+
SWIR
(0.4–2.5
µm) image spectra. Spatial image co-registration of the hy-
perspectral images was performed using homologous points
automatically extracted using the scale-invariant feature
transform (
SIFT
) algorithm. The effect of input image selection
on identifying homologous points through
SIFT
and different
transformation techniques on image co-registration of pan-
oramic ground-based hyperspectral cameras were evaluated.
Regardless of the input images or transformation technique
employed sub-pixel co-registration
RMSE
could be achieved.
Even though the result of this study is satisfactory there is
still room for improvement in accuracy and robustness of the
registration routine. As such, as an alternative to relying on a
single
SIFT
run, combining multiple
SIFT
runs of different im-
age bands may be explored to achieve better spatial distribu-
tion of homologous points across the entire image.
Although proper spatial image co-registration was ex-
pected to provide persistent spectral data, depending upon
intrinsic camera sensor properties and limitations, obtaining
continuous
VNIR
+
SWIR
image spectra from different cameras
may not be straightforward and require characterizing and
correcting the errors and inconsistencies due to these limita-
tions. It has been shown here that a continuous
VNIR
+
SWIR
spectrum is more valuable than either one alone. As such,
using the obtained continuous
VNIR
+
SWIR
image spectra from
separate ground-based hyperspectral cameras provided the
opportunity to differentiate lithological units in the studied
outcrop, which otherwise would not have been possible using
a single camera due to their similar spectral responses.
Acknowledgments
We would like to thank the members of GeoRS Lab at Univer-
sity of Houston: Ms. Diana Krupnik and Dr. Lei Sun for their
help during the fieldwork and Dr. Preston Hartzell of
NCALM
for fruitful discussions. The National Science Foundation
(
NSF
) is acknowledged for funding the purchase of the hyper-
spectral imaging system (Award # 125602). We also like to
thank anonymous reviewers for their constructive criticism.
References
Aguilera, C., F. Barrera, F. Lumbreras, A.D. Sappa, and R. Toledo,
2012. multispectral image feature points,
Sensors
, 12:12661–
12672,
Alonso de Linaje, V., and S.D. Khan, 2017. Mapping of diagenetic
processes in sandstones using imaging spectroscopy:
A case study of the Utrillas Formation, Burgos, Spain,
Sedimentary Geology
, 353:114–124.
sedgeo.2017.03.010
Alonso de Linaje, V., S.D. Khan, and J. Bhattacharya, 2018. Study
of carbonate concretions using imaging spectroscopy in the
Frontier Formation, Wyoming. International Journal of Applied
Earth Observation and Geo-information, 66:82–92.
.
org/10.1016/j.jag.2017.11.010
Table 3. Classification accuracies generated for
VNIR+SWIR
and
SWIR
images.
VNIR+SWIR image
Mudstone (Fe-rich)
Lime Mudstone Hematitic Limestone Mudstone Mudstone (Fe-CO
3
-rich)
Mudstone (Fe-rich)
840
10
11
11
16
Lime Mudstone
0
508
0
2
0
Hematitic Limestone
2
35
810
17
5
Mudstone
0
1
0
516
0
Mudstone (Fe-CO
3
-rich)
1
3
5
0
349
Unclassified
Pixels
20
39
110
2
4
Producer Accuracy
97.33%
85.23%
86.54%
94.16%
93.32%
Overall Accuracy
91.14%
Kappa Coefficient
0.89
SWIR image
Mudstone (Fe-rich)
Lime Mudstone Hematitic Limestone Mudstone Mudstone (Fe-CO
3
-rich)
Mudstone (Fe-rich)
790
9
18
7
0
Lime Mudstone
0
49
1
0
0
Hematitic Limestone
1
308
767
4
7
Mudstone
0
4
0
369
0
Mudstone (Fe-CO
3
-rich)
0
2
3
3
271
Unclassified Pixels
65
245
153
147
82
Producer Accuracy
92.29%
7.94%
81.42%
69.62%
75.28%
Overall Accuracy
67.96%
Kappa Coefficient
0.61
788
December 2018
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
