is the addition of reflectance in blue, green, and red wave-
lengths. Unfortunately, the approach is unable to process opti-
cal data from a single band or bands outside of visible region.
The wavelet transform algorithm decomposes the input image
into images to distinguish clouds and non-cloud features. A
high-pass filter is used to remove the cloud, and a low-pass
filter to enhance the low frequency components affected by
the hi-pass filter. The filtered images are then combined to
complete the process of cloud removal. The transform from
spatial domain to frequency domain and back to spatial
domain could make the algorithm inefficient when the image
size is large. Therefore, an approach using newly available
Band 9 and
QA
band of Landsat-8 to remove cirrus as well as
remaining thin cirro- and alto-clouds is studied.
Analytical Approach
Landsat-8 and Opportunity for Cloud Removal
Landsat-8 was successfully launched on 11 February 2013
(Roy
et al
., 2014). Onboard are the operational land imager
(
OLI
) and thermal infrared sensors (
TIRS
). Compared to previ-
ous Landsat sensors, they are improved in the signal-to-noise
ratio and level of radiometric quantization. Additionally,
Band 9 designed for the qualification of reflectance from cirro-
clouds is available.
QA
band provides information on the pres-
ence/absence of clouds. Therefore, Band 9 and
QA
band offer
the unique opportunity for the removal of thin clouds.
The Algorithm for Removal of Thin Clouds
Bands 1 through 7 of a Landsat-8 image are atmospherically
corrected, and the output image is called as
IMG1
for short.
The algorithm is detailed next.
Removal of Cirrus Clouds Using Band 9
Cirrocumulus, cirrostratus, and cirrus exist normally above 6
km in altitude. The clouds consist of mainly ice crystals. The
reflectance of solar radiation centered at wavelength of 1373
nm is primarily from the ice particles (Gao
et al.
, 1993). Band
9 (1360 - 1390nm) should detect well the presence/absence of
cirrus clouds from space. Therefore, reflectance data of Band
9 are used for the removal of cirrus clouds.
Gao
et al.
(1998) and Gao
et al.
(2002) establish a linear
relationship between the reflectance at 1380 nm (
γ
1380
) and cir-
rus effect on the reflectance of wavelengths from 400 to 1000
nm
(
γ
c
) as
γ
c
=
α
·
γ
1380
(1)
where
α
is a constant. In this study,
γ
1380
is approximated by
the reflectance of Band 9. Thus, after the removal of cirrus
clouds, the surface reflectance,
γ
*
λ
at wavelength of
λ
is:
γ
*
λ
=
γ
λ
–
γ
c
(2)
where
γ
λ
is the reflectance after atmosphere correction at
wavelength
λ
. The removal is separately applied to each band
of Bands 1 through 5. The corrected Bands 1 through 5 is
denoted as
IMG2
.
Removal of Remaining Thin Clouds Using
QA
Band
Bits 14 and 15 of
QA
band indicate the confidence levels of
cloud presence or absence within a Landsat-8 image. In par-
ticular, 00 of bits 14 and 15 means that the status for the cloud
detection cannot be determined (The number of such pixels
in the
QA
band is actually 0.) 01, 10, and 11 indicate that the
confidence levels in cloud detection are 0 to 33 percent, 34 to
66 percent, and 67 to 100 percent, respectively. In this study,
pixels with 01 of bits 14 and 15 are treated as cloud-free. Pix-
els with 10 or 11 are considered as cloud pixels.
For all cloud and cloud-free pixels, reflectance values of
Band 6 (1560 - 1660 nm) and Band 7 (2100 - 2300 nm) are ex-
tracted. Then, measured by reflectance values of both bands,
the distance from the
j
th
cloud pixel to the
k
th
cloud-free pixel,
d
j,k
is calculated as:
d
IMG a b IMG x y
j k
j B
k B
i
i
i
,
,
,
( , )
( , )
=
−
(
)
=
∑
2
2
2
6
7
(
j
= 1, 2, …,
J
;
k
= 1, 2, …,
K
)
(3)
where
IMG
2
j,B
i
(
a,b
)
is the reflectance value of Band 6 (
B
6
) or
Band 7 (
B
7
) of the
j
th
cloud pixel that is delineated by
QA
band
at location (
a
,
b
).
IMG
2
k,B
i
(
x,y
) is the reflectance value of
B
6
or
B
7
of the
k
th
cloud-free pixel at location (
x
,
y
).
J
is the total
number of cloud pixels.
K
is the total number of cloud-free
pixels. For an image with
M
(column)×
N
(row), 1
≤
a
≤
M
, 1
≤
b
≤
N
, 1
≤
x
≤
M
, 1
≤
y
≤
N
, and
J
+
K
=
M
×
N
.
Since solar energy in Bands 6 and 7 can penetrate thin
clouds (Berger
et al.
, 2001), the search program for the mini-
mum value of
d
j,k
values of all cloud-free pixels (
k
= 1, 2, …,
K
) and the
j
th
cloud pixel is:
(
x
'
,y
') = {(
x,y
)|min(
d
j,k
)}.
(4)
The reflectance values of Bands 1 through 5 of the cloud-
free pixel at (
x
'
,y
') are considered as the reflectance values of
Bands 1 through 5 of the
j
th
cloud pixel at (
a
,
b
). Thus, the re-
flectance value from the cloud pixel at (
a
,
b
) in each band is:
TMP1
B
i
(
a,b
) =
IMG
2
B
i
(
a,b
) –
IMG
2
B
i
(
x
'
,y
')
(5)
Thus, thin clouds within each band,
TMP1
B
i
(
i
= 1, 2, 3, 4,
5) are initially obtained. All bands are collectively called as
TMP1
.
To ensure the spatial homogeneous characteristics of
clouds, we apply a smoothing filter to
TMP1
. Then, the fil-
tered result,
TMP2
B
i
(
a,b
) is:
TMP2
B
i
(
a,b
) =
TMP1
B
i
(
a,b
)
7
k
(
p,q
)
(6)
where
7
is the convolution operator.
k
(
p,q
) is the filter
kernel such as a Gaussian filter. The filtering is done for all
cloud pixels of each band. Thus, thin clouds within Bands 1
through 5 are derived, and labeled as
TMP2
. Since the filter-
ing is only performed on the cloud pixels, one obtains not
only satisfactory delineation of thin clouds as shown later,
but also avoids unnecessarily filtering to cloud-free pixels.
Finally, subtracting
TMP2
from
IMG2
, one obtains the cloud
removed product,
IMG3
.
Study Area
The area of research interest is within Sichuan basin, China.
The climate of the basin is humid subtropical. The level of
moisture content in air is high. Due to the basin topography,
the air tends to be stationary. The cloud or fog cover is persis-
tent and extensive. The search at the USGS web site (
http://
earthexplorer.usgs.gov
) indicates that the number of Landsat
images that contain 10 percent or less in cloud cover is very
small even though thousands of Landsat -4, -5, -7, and -8 data
have been acquired so far.
A Landsat-8 image of 129/39 (path/row) acquired on 16
December 2013 was downloaded at
.
gov
. Then, a subimage of 400 × 400 or 12 km × 12 km, cen-
tered at 30°20’4.29”N and 105°3’23.19”E was extracted. The
area was near Ziyang, Sichuan Province. Reflectance values of
Bands 1 through 5 of the subimage were shown in gray scale
(Figure 1a and 1e). The brighter the tone was, the higher the
value was. With concentrated spatial patterns visually, thin
722
September 2015
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
