Removal of Thin Clouds Using Cirrus and
QA Bands of Landsat-8
Yang Shen, Yong Wang, Haitao Lv, and Hong Li
Abstract
After atmospheric correction, an algorithm for the removal
of thin cirrus cloud as well as alto-thin clouds or thin clouds
collectively within visible and near infrared bands (Bands 1
through 5) of Landsat-8 was developed. The algorithm re-
moved cirrus clouds using Band 9 first, and the remaining thin
clouds using quality assurance (QA) band. Using a Landsat-8
sub-image of 129/39 (path/row) acquired on 16 December
2013, we evaluated the algorithm. Thin clouds disappeared
visually. Reflectance values of Bands 1 through 4 decreased
in both steps. Reflectance values of Band 5 decreased in step
one, and then stayed the same. With a nearly cloud-free image
acquired on 30 November 2013 as the “truth,” the spatial cor-
relation coefficients of cloud-covered pixels within the Decem-
ber image were 0.84 or higher. Changes in reflectance values
of Bands 1 to 5, and the high correlation coefficient values
indicated the validity of the algorithm.
Introduction
As solar radiation enters the Earth’s atmosphere, the radiation
in the optical region is selectively scattered and absorbed.
Clouds in the atmosphere further influence the radiation, and
ultimately affect the reflectance of ground targets measured by
an optical remote sensor. The reflectance from the atmosphere
or clouds is not desirable. Both can adversely affect optical
data and data applications. Thus, they need to be removed or
corrected. The atmospheric correction is fruitful. Valid algo-
rithms include atmospheric correction now (
ACORN
) (ImSpec,
L.L.C., 2002), atmospheric correction program (
ATCOR
) (Thie-
mann and Hermann, 2002), atmospheric removal program
(
ATREM
) (CSES, 1999), and fast line-of-sight atmospheric anal-
ysis of spectral hypercubes (
FLAASH
) (Anderson
et al.
, 2002).
On the other hand, the cloud removal in optical data such as
Landsat data remains a challenge. Clouds vary temporally and
spatially. Therefore, a large number of Landsat images can be
contaminated by clouds. Even today, the retrieval of accurate
surface reflectance beneath cloud cover is not operational for
Landsat data.
Clouds appear in the form of cumulus, stratus, or cirrus
from the Earth surface to approximately 12,000 m. Cumulus
clouds are puffy and vertically-developed. Stratus clouds are
flat, layered, and horizontally-developed. Strato or low clouds
(nimbostratus, stratus, and stratocumulus) could be complete-
ly opaque to optical sensors. Cirrocumulus, cirrostratus, and
cirrus are typically thin, and are collectively called as thin
clouds in this study. Because the thin clouds are usually high
in the sky, their exact altitude above the ground are not deriv-
able from satellite images alone. Thus, cloud shadows on the
ground are not considered.
The first step in cloud removal is to identify the cloud. For
Landsat data, numerous identification approaches have been
developed (e.g., Zhu and Woodcock, 2012; Zhu
et al
., 2015).
For instance, the Fmask algorithm has been designed for the
detection of cloud and cloud shadow, as well as snow on
the ground. In addition, indices derived from multispectral
bands, especially thermal infrared bands, are studied to detect
clouds, particularly thick ones. However, the determination
of cloud/non-cloud threshold is somewhat cumbersome. On
the other hand, the quality assessment (
QA
) band of Landsat-8
provides information on the presence or absence of clouds
.
QA
band is used to identify cloud directly here.
Once the cloud is identified, the cloud removal is in order.
Current approaches to remove cirro-clouds as well as thin
alto-clouds can be divided into two groups. The first involves
analyzing multiple images (Chen
et al.
, 2005; Meng
et al.
,
2009). Reflectance from ground that is free of cloud cover
at time one is used to replace the reflectance from the same
ground that is cloud covered at time two. Temporal varia-
tion of datasets in time 1 and time 2 is always a concern
because any phonological change or environmental variation
between two dates could lead errors in cloud corrected data.
The second group of algorithms is based on a single image.
Examples are the homomorphism filter algorithm (Zhao and
Zhu, 1996), the intensity, hue, and saturation (IHS) transform
algorithm (Souza
et al
., 2003), and the wavelet transform al-
gorithm (Guo, 2008). The homomorphism filter algorithm acts
as a high-pass filter because the cloud tends to be uniform
spatially and consists of mainly low-spatial frequency com-
ponents. In the IHS transform approach, the cloud removal
is achieved in the processing of the intensity
component that
Yang Shen is with the School of Resources and Environment,
University of Electronic and Science Technology of China
(UESTC),2006 Xiyuan Avenue, West Hi-tech Zone, Chengdu
City, Sichuan 611731, China
Yong Wang is with the School of Resources and Environment
University of Electronic and Science Technology of China
(UESTC), 2006 Xiyuan Avenue, West Hi-tech Zone, Chengdu
City, Sichuan 611731, China; the Dept. of Geography,
Planning, and Environment, East Carolina University,
Greenville, NC 27858; and the Institute of Remote Sensing
Big Data, Big Data Research Center of UESTC, 2006 Xiyuan
Avenue, West Hi-tech Zone, Chengdu, Sichuan 611731, China
(
).
Haitao Lv is with the School of Resources and Environment
University of Electronic and Science Technology of China
(UESTC), 2006 Xiyuan Avenue, West Hi-tech Zone, Chengdu
City, Sichuan 611731, China.
Hong Li is with the Dept. of Management and Information
Systems, East Carolina University, Greenville, NC 27858.
Photogrammetric Engineering & Remote Sensing
Vol. 81, No. 9, September 2015, pp. 721–731.
0099-1112/15/721–731
© 2015 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.81.9.721
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2015
721
