sensor artifact removal, and atmospheric compensation and
reflectance conversion (Kurz
et al
., 2013; Okyay
et al
., 2016).
For this purpose, in-house preprocessing routines were de-
veloped using
MATLAB
2016b (The Mathworks, Inc.). The dark
current signal was removed using an average frame obtained
from the dark image collected during data acquisition. Along
with the manufacturer provided non-uniformity correction
(
NUC
), dark frame subtraction removed most of the systematic
striping in the imagery. Image artifacts associated with the
poor performance of individual elements in the 2D sensor ar-
ray and push-broom scanning were identified on each image
band and replaced by the mean of adjacent lines in the spatial
domain. Brightness gradient caused mainly by the cylindri-
cal geometry of the panoramic images were corrected for each
spectral band as outlined in Okyay
et al
. (2016). For atmo-
spheric compensation and reflectance conversion, Empirical
Line Calibration (
ELC
) (Smith and Milton, 1999) was used
employing two calibration panels with ~3% and ~99% re-
flectance. Some of the image artifacts caused by bad pixels on
the sensor array can only be seen after the reflectance conver-
sion. Thus, following the reflectance conversion, bad pixels
were identified in the spectral domain and corrected using
linear interpolation. Subsequently, prior to spatial image co-
registration the
VNIR
image was down-sampled, i.e., the size of
the
VNIR
image was decreased, to match the dimensions of the
SWIR
image.
Spatial Image Co-registration
The homologous points between the ground-based hyperspec-
tral images needed for spatial co-registration were automati-
cally extracted using a feature-based image matching algo-
rithm: Scale-invariant Feature Transform (
SIFT
) (Lowe, 2004).
The implementation of
SIFT
used in this study (Wu, 2007) is
derived from the original work of Lowe (2004). The
SIFT
al-
gorithm has been widely used in remote sensing applications
due to its robustness to variable geometric and radiometric
conditions (Khan
et al
., 2011; Wessel
et al
., 2007). Although
a brief overview of the
SIFT
algorithm is provided, it is by no
means an exhaustive discussion. Thus, for further details of
the
SIFT
algorithm and parameters the reader is referred to the
original work of Lowe (2004 and 1999) which is considered to
be the best material on
SIFT
.
SIFT
initially identifies potential interest points (keypoints)
on input images separately. This is achieved by searching
local extrema over Difference of Gaussians (
DoG
) obtained
by differencing adjacent images progressively blurred with
Gaussians. However, not all keypoints will be stable such as
those have low contrast or are poorly localized (i.e., along
edges). If the intensity of a keypoint is less than a thresh-
old, called
contrast threshold
, it is considered as potentially
unstable and rejected. Similarly, if the ratio between the
principal curvatures of a keypoint is greater than a threshold,
called
edge threshold
, it is considered poorly localized and
rejected. While these parameters are rather arbitrary, as per
Lowe (2004), a
contrast threshold
of 0.03, an
edge thresh-
old
of 10, and a
nearest neighbor ratio
of 0.8 were used in
this study. Each stable keypoint is then assigned a principal
orientation based on local image gradient directions. Subse-
quently, an invariant keypoint descriptor which is essentially
a spatial histogram based on image gradient magnitude and
orientation and represented as a 128-dimensional feature
vector is created at each keypoint. Following their deriva-
tion, distinctive keypoints in individual images are matched
based on minimum Euclidean distance between the invariant
descriptors (i.e., Nearest Neighbor search). However, not all
keypoints will have correct matches and thus, the keypoints
with incorrect matches need to be discarded. For this, ratio of
distances to closest match and second-closest match, called
nearest neighbor ratio
is calculated. If the ratio is greater than
a threshold, the match is considered incorrect and eliminated.
Lowe (2004) has shown that a
nearest neighbor ratio
thresh-
old at 0.8 eliminates 90% of incorrect matches while elimi-
nating only less than 5% of correct matches. Any outliers, i.e.,
incorrect matches, which may still remain within matching
points identified in
SIFT
was further eliminated using the Ran-
dom Sampling and Consensus (
RANSAC
) algorithm (Fischler
and Bolles, 1981) leaving only the inlier points.
RANSAC
is an
iterative algorithm to estimate a mathematical model from a
data set containing outliers. As
RANSAC
works by identifying
the outliers in a data set, it could also be interpreted as an
outlier detection method (Strutz, 2011).
Although the
SIFT
algorithm is robust to radiometric
changes, it could further benefit from the use of images
acquired at similar wavelengths like any other feature-based
matching algorithm (Schwind
et al
., 2014; Sima and Buck-
ley, 2013). However, the low signal-to-noise ratio of the
VNIR
image bands within the overlapping spectral range (briefly
discussed below) causes speckles in the images, which sub-
sequently affect keypoint identification and image matching.
In attempt to evaluate the effect of input image selection on
image matching, four independent runs of
SIFT
were per-
formed using the same parameters but different input images.
Once the homologous points were extracted, the spatial image
co-registration was performed in ENVI version 5.3 (Harris
Geospatial, Boulder, Colorado) using both affine and first-
order polynomial transformations. The affine transformation
accounts for rotation, scaling, and translation (
RST
) but does
not allow for shearing, whereas the polynomial transforma-
tion accommodates image shearing in addition to rotation,
scaling, and translation (ENVI User Guide, 2017). Thus, in
order to account for potential shearing in the images, first-
order polynomial transformation is used in addition to affine
transformation.
Spectral Concatenation
Proper spatial co-registration of the ground-based hyperspec-
tral images, in theory, should provide continuous
VNIR
+
SWIR
spectra that allow a more complete spectral analysis. Unfor-
tunately, in reality this is an overly optimistic expectancy
due mainly to non-optimal sensor properties. Dissimilar
spectral performance of the hyperspectral cameras hampers
the spectral concatenation of
VNIR
and
SWIR
images. A pre-
liminary analysis of image spectra showed a large increase
in noise towards the longer end of the
VNIR
camera spectrum
caused by decreasing sensitivity of the camera sensor (Figure
2). The noise in wavelengths longer than approximately 920
nm becomes more prominent and potentially preventative
for spectral analysis (
in Figure 2); therefore, image bands
of these wavelengths were excluded from further analysis (
in Figure 2). In addition, due to the relatively poor spectral
Table 1. Specifications of the
VNIR
and
SWIR
cameras from
SpecIm, Finland.
VNIR Camera
SWIR Camera
Spectrograph ImSpector V10E
ImSpector N25E
Spectral range
400 – 1000 nm
900 – 2500 nm
Spectral Sampling 0.72 – 5.8 nm
1
6.3 nm
Sensor
HgCdTe (MCT)
Charged-couple
device (CCD)
Spatial dimension up to 1600 pixels
2
320 pixels
Spectral dimension up to 840 pixels
1
256 pixels
Pixel pitch
7.4 µm
30 µm
Digitization
12-bit
14-bit
Frame rate
up to 120 fps
up to 100 fps
1
Adjustable by spectral binning;
2
Adjustable by spatial binning
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
December 2018
783
