ChapterPDF Available

Cloud and Cloud Shadow Detection for Landsat Images: The Fundamental Basis for Analyzing Landsat Time Series

Authors:
3
1
Cloud and Cloud Shadow Detection for
Landsat Images: The Fundamental Basis
for Analyzing Landsat Time Series
Zhe Zhu, Shi Qiu, Binbin He, and Chengbin Deng
CONTENTS
Brief Summary ......................................................................................................... 4
1.1 Introduction .................................................................................................... 4
1.2 Landsat Data and Reference Masks ............................................................ 5
1.2.1 Landsat Data .......................................................................................5
1.2.2 Manual Masks of Landsat Cloud and Cloud Shadow ..................7
1.3 Cloud and Cloud Shadow Detection Based on a Single-Date
Landsat Image ................................................................................................ 8
1.3.1 Physical-Rules-Based Cloud and Cloud Shadow Detection ........ 8
1.3.1.1 Physical-Rules-Based Cloud Detection Algorithms ...... 8
1.3.1.2 Physical-Rules-Based Cloud Shadow Detection
Algorithms ......................................................................... 12
1.3.2 Machine-Learning-Based Cloud and Cloud Shadow
Detection ........................................................................................... 14
1.4 Cloud and Cloud Shadow Detection Based on Multitemporal
Landsat Images ............................................................................................ 14
1.4.1 Cloud Detection Based on Multitemporal Landsat Images ...... 15
1.4.2 Cloud Shadow Detection Based on Multitemporal Landsat
Images ................................................................................................16
1.5 Discussions ................................................................................................... 17
1.5.1 Comparison of Different Algorithms ........................................... 17
1.5.2 Challenges ......................................................................................... 17
1.5.3 Future Development ........................................................................ 18
1.5.3.1 Spatial Information ........................................................... 18
1.5.3.2 Temporal Frequency ......................................................... 18
1.5.3.3 Haze/Thin Cloud Removal ............................................. 18
1.6 Conclusion ....................................................................................................19
References .............................................................................................................. 19
4Remote Sensing Time Series Image Processing
Brief Summary
Cloud and cloud shadow detection is the fundamental basis for analyzing
Landsat time series. This chapter provides a comprehensive review of all the
cloud and cloud shadow detection algorithms designed explicitly for Landsat
images. This review provides guidance on the selection of cloud and cloud
shadow detection algorithms for various applications using Landsat time
series.
1.1 Introduction
Landsat satellites have been widely used for a variety of remote sensing
applications, such as change detection (Collins and Woodcock, 1996; Xian
etal., 2009), land cover classication (Homer etal., 2004; Yuan etal., 2005),
biomass estimation (Zheng etal., 2004; Lu, 2005), and leaf area index retrieval
(Chen and Cihlar, 1996; Fassnacht etal., 1997). Nevertheless, for decades, most
of the analyses were based on a single or a few cloud free Landsat images
acquired at different dates, due to the high cost of Landsat images prior to
2008 (Loveland and Dwyer, 2012). Free and open access to the entire Landsat
archive in 2008 has changed the story entirely (Woodcock etal., 2008; Wulder
etal., 2012). Landsat data are being downloaded for an unprecedented variety
of applications. Many of them require frequent Landsat observations for
the same location – Landsat Time Series (LTS). The Landsat Global Archive
Consolidation (LGAC) initiative has added 3.2 million Landsat images to the
U.S. Geological Survey (USGS) Earth Resources Observation and Science
(EROS) Center (Wulder etal., 2016), which has made time series analysis
with LTS even more popular. Decreasing data storage costs and increasing
computing power have further stimulated the use of LTS.
Though time series analysis based on LTS has attracted much attention,
automated cloud and cloud shadow detection has been and remains a major
obstacle. The presence of clouds and cloud shadows reduces the usability of
the Landsat image which makes it difcult for any kind of remote sensing
applications. For coarse resolution images, such as from the Advanced Very
High Resolution Radiometer (AVHRR) and Moderate Resolution Imaging
Spectroradiometer (MODIS), there are many mature operational algorithms
for detecting clouds and cloud shadows (Derrien etal., 1993; Ackerman
etal., 1998). However, for moderate resolution satellites, like Landsat, there
were no algorithms that could provide cloud and cloud shadow masks at
the pixel level. This is not surprising because Landsat images were not
affordable, each of which previously cost more than 400 U.S. dollars per
image. Even when cloudy Landsat images are used, most of the time only
5Cloud and Cloud Shadow Detection for Landsat Images
a small number of images are needed, and manual interpretation of clouds
and their shadows in the images is feasible. However, when these nancial
constraints were lifted (Woodcock etal., 2008), an unprecedented demand
arose for automatically processing a massive number of Landsat images for
time series analysis. Manual interpretation of cloud and cloud shadow was
no longer acceptable.
1.2 Landsat Data and Reference Masks
1.2.1 Landsat Data
Since 1972, Landsat satellites have provided a continuous Earth observation
data record. Landsats 1–5 carried the Multispectral Scanner System (MSS)
sensor with 60-meter spatial resolution. The MSS only collected images with
four spectral bands, including green, red, and two Near InfraRed (NIR)
bands (Ta ble 1.1). Note that the Landsat 3 MSS also included a Thermal
Infrared (TIR) band, but failed shortly after launch. The fewer bands result
in known difculties in detecting clouds and cloud shadows (Braaten etal.,
2015). However, the MSS images are still crucial for LTS related analyses
(Pugmacher etal., 2012). Since the launch of Landsat 4 in 1982, the Thematic
Mapper (TM) has provided more spectral information at 30-meter spatial
resolution (Table 1.1). The TM sensor was also carried on Landsat 5, which
was launched on March 1, 1984, and functioned for over 28 years until 2012.
Landsat 7, carrying the Enhanced Thematic Mapper Plus (ETM+), was
launched on April 15, 1999 (Table 1.1). This instrument also has a 30-meter
spatial resolution and improved radiometric and geometric calibration
accuracies, but the Scan Line Corrector (SLC) has failed since May 31, 2003.
Both TM and ETM+ have a TIR band at a spatial resolution of 120-meter
and 60-meter, respectively. Landsat 8 was launched on February 11, 2013.
It has two sensors: Operational Land Imager (OLI) and Thermal Infrared
Sensor (TIRS) (Tabl e 1.1). The OLI instrument provides 30-meter resolution
optical data, while TIRS provides 100-meter resolution TIR data. Note that
the TIRS has a shorter design life compared to the OLI. Additionally, the new
OLI added the new blue band (Band 1: 0.435–0.451 µm) and the cirrus band
(Band9: 1.363–1.384 µm) with 30-meter spatial resolution.
Although each Landsat satellite can cover global land every 16 days, many
of the observations are inevitably impacted by clouds and cloud shadows.
Figure 1.1 illustrates mean global cloud cover calculated based on all available
Landsat 8 daytime images acquired between September 2013 and August
2017. The cloud cover information for each Landsat Path/Row is calculated
based on the metadata of Landsat 8 images downloaded from the USGS
Landsat Bulk Metadata Service (https://landsat.usgs.gov/landsat-bulk-
metadata-service), which is derived based on an algorithm called Fmask
6Remote Sensing Time Series Image Processing
(Zhuand Woodcock, 2012; Zhu etal., 2015). Extremely high cloud cover is
observed in tropical rainforest regions, while for arid places, such as desert
or dryland regions, cloud cover is relatively low. The mean global cloud cover
contained in the Landsat images is approximately 41.59%, which means that
clouds impact almost half of the Landsat observations.
TABLE 1.1
Landsat 1–5 MSS, Landsat 4–5 TM, Landsat 7 ETM+ and Landsat 8 OLI Sensor
Characteristics
Landsat 1–5
MSS Bands (µm)
Landsat 4–5
TM Bands (µm)
Landsat 7
ETM+ Bands (µm)
Landsat 8
OLI/TIRS Bands (µm)
Band 1 (0.45–0.52) Band 1 (0.45–0.52) Band 1 (0.435–0.451)
Band 4 (0.50–0.60) Band 2 (0.52–0.60) Band 2 (0.52–0.60) Band 2 (0.452–0.512)
Band 5 (0.60–0.70) Band 3 (0.63–0.69) Band 3 (0.63–0.69) Band 3 (0.533–0.590)
Band 6 (0.70–0.80)
Band 7 (0.80–1.10)
Band 4 (0.76–0.90) Band 4 (0.77–0.90) Band 4 (0.636–0.673)
Band 5 (1.55–1.75) Band 5 (1.55–1.75) Band 5 (0.851–0.879)
Band 8 (10.40–12.50)
Landsat 3 onlya
Band 6 (10.40–12.50) Band 6 (10.40–12.50) Band 6 (1.566–1.651)
Band 7 (2.08–2.35) Band 7 (2.09–2.35) Band 7 (2.107–2.294)
Band 8 (0.52–0.90) Band 8 (0.503–0.676)
Band 9 (1.363–1.384)
Band 10 (10.60–11.19)
Band 11 (11.50–12.51)
a Indicates that the thermal band of the Landsat 3 MSS was unsuccessful and not available.
Mean cloud cover percentage for each scene (%)
0–20 21–40 41–60 61–80 81–100 No data
FIGURE 1.1
Mean global cloud cover percentage calcu lated based on all available La ndsat 8 images acquired
between September 2013 and August 2017. A total of 966,708 Landsat 8 images are used. The
mean global cloud cover percentage from all Landsat 8 observations is 41.59%.
7Cloud and Cloud Shadow Detection for Landsat Images
1.2.2 Manual Masks of Landsat Cloud and Cloud Shadow
Manually interpreted cloud and cloud shadow masks are the most important
data source for developing and/or validating the cloud and cloud shadow
detection algorithms (Irish et al., 2006; Zhu and Woodcock, 2012; Hughes
and Hayes, 2014; Foga etal., 2017; Qiu etal., 2017). At present, there are
three publicly available, manually interpreted cloud and cloud shadow
masks derived from Landsat images (Tabl e 1. 2), including “L7 Irish” masks
for Landsat 7 data (USGS, 2016a), “L8 SPARCS” masks for Landsat 8 data
(USGS, 2016b), and “L8 Biome” masks for Landsat 8 data (USGS, 2016c). These
masks are manually interpreted based on Landsat images randomly selected
from different locations, which cover a variety of land cover types, and the
cloud cover percentage within each manual mask also varies substantially.
The “L7 Irish” manual masks were rst created to systematically coverthe
global environments and different cloud conditions (Irish etal., 2006). The
“L7 Irish” masks were produced based on Landsat 7 ETM+ images by
visual interpretation of full resolution images with different Landsat band
combinations, and their average error was estimated at approximately 7%
(Oreopoulos et al., 2011). The “L8 SPARCS” manual masks were created
manually from Landsat 8 OLI images by Hughes and Hayes (2014), which
was used to validate Spatial Procedures for Automated Removal of Cloud
and Shadow (SPARCS) algorithm. Note that those manual cloud and cloud
shadow masks are provided at a 3 km by 3 km Landsat subset (1000 × 1000
30-meter pixels), with around 4% of pixels being ambiguous (Foga etal., 2017).
The manual cloud and cloud shadow masks in “L8 Biome” are designed for
Landsat 8 OLI/TIRS images, which were randomly selected from different
locations around the world using a biome-based stratied sampling approach.
Their corresponding cloud and cloud shadow masks were produced by
multiple visual criteria (such as brightness, shape, and texture) with various
band combinations by a single analyst (Foga etal., 2017). This new dataset
achieved better accuracy than the “L7 Irish,” due to the multiple visual
criteria it used (Foga etal., 2017).
TABLE 1.2
Manual Cloud and Cloud Shadow Masks Derived from Landsat Images
Name Sensor
Number
of
Images
Date Range
Error ReferenceStart End
L7 Irish ETM+
(SLC on)
206 (45) 06/06/2000 12/30/2001 7.00% USGS (2016a)
L8 SPARCS OLI 80 (80) 05/12/2013 11/02/2014 4.00% USGS (2016b)
L8 Biome OLI 96 (33) 04/13/2013 11/05/2014 Less than
7.00%
USGS (2016c)
Note that the all images contain manual cloud masks. The numbers in the brackets indicate the
number of cloud shadow masks for each dataset.
8Remote Sensing Time Series Image Processing
1.3 Cloud and Cloud Shadow Detection Based
on a Single-Date Landsat Image
Recently, many cloud and cloud shadow detection algorithms have been
developed for Landsat images (Table 1. 3). Among them, some were proposed
by using a single-date Landsat image (hereafter single-date algorithms), and
we can classify these single-date algorithms into two categories: physical-
rules-based and machine-learning-based algorithms (Table 1.3).
1.3.1 Physical-Rules-Based Cloud and Cloud Shadow Detection
1.3.1.1 Physical-Rules-Based Cloud Detection Algorithms
The physical-rules-based algorithms detect clouds by identifying their
physical characteristics of clouds, that are “bright”, “white”, “cold, and
“high” (Irish, 2000; Zhu etal., 2015). Compared to other land cover types, the
reectance of cloud is much higher in almost all wavelengths, which makes
clouds look “bright”. Therefore, we can use some simple thresholds in the
spectral bands to exclude clear sky pixels that are not bright enough. Clouds
are “white” due to the similar reectance in all wavelengths, particularly
in the visible bands. In this case, some indices such as “whiteness” (Zhu
and Woodcock, 2012), Normalized Difference Vegetation Index (NDVI), and
Normalized Difference Snow Index (NDSI) can be used to separate clouds
from clear sky pixels that are not white enough. Moreover, clouds are “cold
because they are usually high in the air, and the temperature of clouds
follows the environmental lapse rate—the higher the clouds, the colder the
temperature. This characteristic can be successfully captured by the thermal
band from Landsat TM, ETM+, and TIRS instruments, which can further
separate clouds from similar bright and white land surfaces (e.g., barren sand,
soil, rock, snow/ice, etc.). Additionally, as clouds are usually “high” in the
sky, the path for water vapor over clouds is much shorter than that for places
without clouds. Therefore, the water vapor absorption band (or the cirrus
band) is especially helpful in identifying higher altitude clouds.
Most of the algorithms are developed for Landsat TM and ETM+ images.
Historically, the Automated Cloud Cover Assessment (ACCA) was used to
provide cloud cover percentage in Landsat TM and ETM+ images (Irish, 2000;
Irish etal., 2006). With several spectral lters, ACCA works well for estimating
a cloud cover score for each image but is not sufciently precise in identifying
the locations and boundaries of clouds (Zhu and Woodcock, 2012). Besides,
ACCA fails to identify warm cirrus clouds and may misidentify snow/ice
as clouds, mainly because the static thresholds in ACCA are insufcient to
capture the various kinds of clouds and the variety of land surface types.
To better distinguish cloud from snow/ice, Choi and Bindschadler (2004)
used the cloud and cloud shadow geometry matching approach iteratively
9Cloud and Cloud Shadow Detection for Landsat Images
TABLE 1.3
Characteristics of Different Cloud and Cloud Shadow Detection Algorithms for Landsat Images
Category
Algorithm
Name Landsat Sensor
Cloud/
Shadow Ancillary Data
Overall
Accuracy Reference
Single-date:
Physical rules
MFmask TM ETM+ OLI/TIRS Both DEM 96% Qiu etal. (2017)
LSR 8 OLI/TIRS Both N/A N/A Vermote etal. (2016)
UDTCDA OLI/TIRS Cloud MOD09A1 N/A Sun etal. (2016)
MSScvm MSS Both DEM 84% Braaten etal. (2015)
ELTK OLI/TIRS Cloud N/A N/A Wilson and Oreopoulos
(2013)
Fmask TM ETM+ OLI/TIRS Both N/A 96% Zhu and Woodcock (2012),
Zhu etal. (2015)
N/A TM ETM+Both DEM 88%99% Huang etal. (2010)
LTK TM ETM+ OLI/TIRS Cloud N/A 93% Oreopoulos etal. (2011)
LEDAPS TM ETM+Both Air temperature
from NCEP
N/A Vermote and Saleous (2007)
CDSM/ANTD ETM+Both N/A N/A Choi and Bindschadler (2004)
ACCA TM ETM+ OLI/TIRS Cloud N/A N/A Irish (2000), Irish etal. (2006)
Single-date:
Machine
learning
N/A OLI Cloud N/A N/A Zhou etal. (2016)
SPARCS ETM+Both N/A 99% Hughes and Hayes (2014)
See5 OLI Cloud N/A 89% Scaramuzza etal. (2012)
AT-ACCA OLI Cloud N/A 90% Scaramuzza etal. (2012)
N/A ETM+Both N/A N/A Potapov etal. (2011)
N/A ETM+Cloud N/A N/A Roy etal. (2010)
N/A MSS Cloud N/A 93% Lee etal. (1990)
(Continued)
10 Remote Sensing Time Series Image Processing
TABLE 1.3 (Continued)
Characteristics of Different Cloud and Cloud Shadow Detection Algorithms for Landsat Images
Category
Algorithm
Name Landsat Sensor
Cloud/
Shadow Ancillary Data
Overall
Accuracy Reference
Multi-date IHOT MSS TM ETM+Cloud N/A N/A Chen etal. (2015)
TmaskaTM ETM+ OLI/TIRS Both N/A N/A Zhu and Woodcock (2014)
N/AaTM ETM+Both N/A 97% Goodwin etal. (2013)
N/A TM ETM+Both N/A N/A Jin etal. (2013)
MTCDaTM ETM+Both Sentinel-2 data N/A Hagolle etal. (2010)
N/A TM Both N/A N/A Wang etal. (1999)
Note: MFmask: Mountainous Fmask; LSR 8: Landsat 8 Surface Reectance product; UDTCDA: Universal Dynamic Threshold Cloud Detection Algorithm;
MSScvm: MSS clear-view-mask; ELTK: Enhanced LTK; Fmask: Function of mask; LTK: Luo Trishchenko Khlopenkov; LEDAPS: Landsat Ecosystem
Disturbance Adaptive Processing System; CDSM/ANTD: Cloud Detection using Shadow Matching/Automatic NDSI Threshold Decision; ACCA:
Automatic Cloud Cover Assessment; SPARCS: Spatial Procedures for Automated Removal of Cloud and Shadow; See5: C5.0 algorithm used to
generate a decision tree. AT-ACCA: Articial Thermal-Automated Cloud Cover Algorithm; IHOT: Iterative Haze Optimized Transformation;
Tmask: multiTemporal mask; MTCD: Multi-Temporal Cloud Detection; TM: Thematic Mapper; ETM+: Enhanced Thematic Mapper Plus; OLI:
Operational Land Imager; TIRS: Thermal Infrared Sensor; DEM: Digital Elevation Model; MOD09A1: MODerate resolution Imaging
Spectroradiometer surface reectance product; NCEP: National Centers for Environmental Prediction.
a Indicates the algorithms based on Landsat time series.
11Cloud and Cloud Shadow Detection for Landsat Images
to determine the optimal threshold of NDSI for each Landsat image in cloud
detection. This approach works well over ice sheets, but it is time-consuming
and only works on the surface of ice sheets. Vermote and Saleous (2007)
proposed a cloud detection algorithm for Landsat TM and ETM+ images,
and the detection results are provided as one of the internal products in
the Landsat Ecosystem Disturbance Adaptive Processing System (LEDAPS)
atmosphere correction software. This algorithm needs surface temperature
from the National Centers for Environmental Prediction (NCEP) as ancillary
data to generate a surface temperature reference layer for cloud detection.
Huang et al. (2010) constructed a spectral temperature space to identify
clouds in Landsat image using clear sky forest pixels as a reference. This
method works well over forest areas but has not been fully tested for non-
forest areas. By revisiting the Luo Trishchenko Khlopenkov (LTK) scene
identication algorithm initially developed for the MODIS image (Luo etal.,
2008), Oreopoulos etal. (2011) modied this algorithm to detect clouds in
Landsat 7 ETM+ data using simple thresholds derived for the blue, red, NIR,
and Short-Wave Infrared (SWIR) bands (no thermal band). Recently, Zhu
and Woodcock (2012) developed the Fmask (Function of mask) algorithm
that detects cloud by using a scene-based threshold. This method is suitable
for the Landsats 4–8 data and can generate a cloud probability layer. Users
can adjust the threshold of cloud probability to determine cloud masks. The
default threshold (global optimal) is 22.5%. If large omissions are found, a
smaller threshold (e.g., 12.5%) is recommended, and if large commissions are
observed, a higher threshold (e.g., 50%) is recommended. This method has
also been successfully integrated into the Landsat surface reectance Climate
Data Record (CDR) and Collection 1 Quality Assessment (QA) band provided
by the USGS Earth Resources Observation and Science (EROS) Center. In
the Fmask algorithm, the thermal band is one of the important inputs, as
it can capture the “cold” character of clouds (Zhu etal., 2015). However, the
temperature for clear sky pixels can also vary widely due to substantial
changes in elevation, and this will lead to commission and omission errors
in cloud detection in mountainous areas. To reduce this issue, Qiu etal. (2017)
provided a Mountainous Fmask (MFmask) algorithm that normalizes the
thermal band with Digital Elevation Models (DEMs) based on a simple linear
temperature-elevation model.
There are also algorithms explicitly designed for Landsat 8 images, many of
which take advantage of the new blue and cirrus bands equipped in Landsat
OLI. Wilson and Oreopoulos (2013) further modied the aforementioned LTK
algorithm by including the cirrus band to detect cloud better. Zhu etal. (2015)
also designed a cloud detection algorithm for Landsat 8 images by calculating
a thin cloud probability layer from the cirrus band, and achieved better
accuracy than the Fmask algorithm designed initially for TM and ETM+
images. Vermote etal. (2016) proposed a new cloud detection algorithm for
Landsat 8, which used the inversion “residual” from the two blue bands and
the cirrus band reectance. To minimize the inuences of cloud detection
12 Remote Sensing Time Series Image Processing
from mixed pixels, complex surface structures, and atmospheric factors,
Sun etal. (2016) presented a Universal Dynamic Threshold Cloud Detection
Algorithm (UDTCDA) for Landsat 8 OLI images, but only the blue, green, red,
NIR, and SWIR bands were used. The dynamic threshold in this method was
determined based on MODIS monthly surface reectance database, which
was established based on the long-time series of MODIS 8-day synthetic
surface reectance products.
Very few algorithms have been designed for the Landsat MSS image, due to
the limited number of spectral bands within the MSS sensor. To address this
issue, Braaten etal. (2015) proposed a simple and automated cloud detection
algorithm relying on green band brightness and the normalized difference
between the green and red bands and achieved comparable accuracies to the
Fmask algorithm.
1.3.1.2 Physical-Rules-Based Cloud Shadow Detection Algorithms
Detecting cloud shadows for Landsat images is more difcult than detecting
clouds due to the spectral similarity of cloud shadows and dark surfaces.
Cloud shadows are usually detected based on physical rules derived from
the cloud shadow geometry.
Previously, the cloud shadow detection algorithms were developed
based on simple spectral tests according to the dark features. However, it
is difcult to directly use thresholds to determine cloud shadows because
their spectral signatures are very similar to other dark surfaces (e.g., terrain
shadows, wetlands, dark urban, etc.). Fortunately, the geometry-based cloud
shadow detection has shown relatively good results. The geometry-based
cloud shadow detection approach is based on the projection of cloud object
onto the local plane of the Earth with respect to a direction of incoming
solar radiation (Berendes et al., 1992; Le Hégarat-Mascle and André,
2009; Simpson etal., 2000). The relative positions of the sun, the satellite,
and the cloud can be used to predict the cloud shadow observed in the
satellite images (Fig u re 1.2). Methods for detecting cloud shadow based on
geometry can be grouped into two categories: shape-similarity-match and
cloud-height-estimation.
The shape-similarity-match approach detects cloud shadow by matching
cloud shadows with cloud objects, assuming that cloud and cloud shadow
shape are similar (Gurney, 1982; Berendes etal., 1992). Gurney (1982) assumed
that a cumulus cloud is approximated in shape by its associated shadow and
matched cloud shadows with clouds. Berendes etal. (1992) developed a semi-
automated methodology for estimating cumulus cloud base height using
Landsat data by matching cloud edges with their corresponding shadow
edges. Due to the absence of the thermal band, Braaten etal. (2015) used cloud
projection to identify cloud shadow in Landsat MSS image based on their
geometry information. Although the computation of cloud and cloud shadow
match is time-consuming and may result in some mismatches, this approach
13Cloud and Cloud Shadow Detection for Landsat Images
is still an attractive routine, especially for images without the thermal band
(e.g., Landsat MSS).
The cloud-height-estimation method uses a constant lapse rate to estimate
cloud top height based on the Brightness Temperature (BT) difference between
cloud top and the land surface. When the cloud height is known, the clouds
can be easily projected to predict their associate shadows on the ground
based on their geometry relationship (Vermote and Saleous, 2007; Huang
etal., 2010). Vermote and Saleous (2007) detected cloud shadow for Landsat
data using a geometric determination of shadow based on the cloud mask and
the estimated altitude of cloud derived from the BT and a conversion factor
range. Huang etal. (2010) identied cloud shadow based on the dark spectral
features, the temperature-based cloud height estimation with a constant lapse
rate, and the solar illumination geometry. Those methods can work well for
thick clouds but are less ideal for the semitransparent clouds, of which the BT
will be a mixture of thin cloud and the surface underneath.
Zhu and Woodcock (2012) and Zhu etal. (2015) calculated cloud shadows
by combining the previous shape-similarity-matching and the cloud-height-
estimation methods, and treated the cloud as a 3D object. This algorithm works
well for most cases, but it may fail to detect the correct cloud shadows for places
of large topographic change and terrain shadows. To address those issues, Qiu
etal. (2017) applied a double-projection method to calculate cloud shadow
shape and removed terrain shadows using a topographic correction model
with the aid of DEMs. In addition, this improved Fmask algorithm can estimate
a more accurate cloud height based on heights estimated from neighboring
clouds, that also improves the detection of cloud shadow for Landsat images.
North
Scan direction
of the sensor
Cloud shadow
Sensor
Cloud
Sunlight direction
h
ϕS
θS
θυ
Observed cloud
Sun
View direction
FIGURE 1.2
Sun/cloud/shadow geometry in Landsat image. Note that θs is the solar zenith angle, ϕs is the
solar azimuth angle, θv is the satellite view zenith angle, and h is the cloud height. The position
of the real cloud may shift from the cloud directly observed by the Landsat sensor.
14 Remote Sensing Time Series Image Processing
1.3.2 Machine-Learning-Based Cloud and Cloud Shadow Detection
Regarding the machine-learning-based algorithms, clouds and/or cloud
shadows are generally treated as a cover type and are identied using
a certain classier trained by previously collected training dataset –
supervised classication. Lee etal. (1990) combined texture-based features
in a network to discriminate clouds in Landsat MSS images and achieved an
overall cloud identication accuracy of 93%. Recognizing that omission and
commission errors for cloud detection will always occur in large datasets
for ACCA, Roy etal. (2010) implemented both the ACCA algorithm and a
classication tree approach to detect clouds using a large number of training
pixels from a global Landsat Level 1G database. Potapov etal. (2011) also
manually selected lots of cloud pixels as training data based on 21 Landsat
images from different years and different regions and built a single tree
model for cloud type to classify the clouds. Due to the possible loss of the
thermal band on Landsat 8, Scaramuzza etal. (2012) expanded the ACCA
pass-1 algorithm without the use of the thermal band, identifying clouds for
Landsat 8 through a statistical classier C5.0 (a classication tree) based on
many randomly sampled pixels from a series of training images. Hughes and
Hayes (2014) also explored the inclusion of spatial information as an input to
a neural network classier on identifying and classifying clouds for Landsat
images. Zhou etal. (2016) utilized the traditional threshold to obtain a coarse
cloud mask and then used the Support Vector Machine (SVM) classier to
detect clouds in Landsat 8 images. Though all these investigations pointed
out the usefulness of machine-learning-based methods in cloud detection,
most require a certain level of knowledge of cloud or surface conditions
within the images (as training data) and commonly fail to detect clouds for
certain unique conditions (Huang etal. 2010). Additionally, cloud shadow
in Landsat image can also be detected using the machine-learning-based
methods (Potapov et al., 2011; Hughes and Hayes, 2014). This method,
however, heavily relies on the training dataset and has substantial omission
or commission errors (Hughes and Hayes, 2014).
1.4 Cloud and Cloud Shadow Detection Based
on Multitemporal Landsat Images
In addition to the single-date algorithm, cloud and cloud shadow detection
algorithms based on multitemporal Landsat images have also been developed
(hereafter multitemporal algorithm). Compared to the spectral or spatial
features derived from a single-date Landsat image, multitemporal Landsat
images can provide extra-temporal information in cloud and cloud shadow
detection, and are reported to produce better cloud and cloud shadow masks
15Cloud and Cloud Shadow Detection for Landsat Images
(Table 1.3). The basic idea of these algorithms is that clouds and cloud shadows
can be easily detected by comparing an observed image with a clear sky
reference (image differencing), as the presence of clouds and cloud shadows
will result in sudden changes of Landsat reectance (Wang etal., 1999; Jin
etal., 2013).
1.4.1 Cloud Detection Based on Multitemporal Landsat Images
For a long time, multitemporal cloud detection algorithms were only based on
two-date or multi-date Landsat images. Early on, Wang etal. (1999) proposed
the use of two-date Landsat TM images to nd clouds by image differencing.
This proposed method rst coarsely nds the clouds for the two Landsat TM
images by setting a histogram-derived threshold for the brightness values,
and then uses another static threshold for the absolute brightness difference
between the two images to further ensure reliable cloud detection. Jin etal.
(2013) identied clouds by incorporating Landsat blue and thermal bands
from two-date images. Based on two Landsat images that have no overlapping
clouds, this method rst selected the relaxed clouds by differencing the blue
bands from the two images and then produced the restricted clouds by
eliminating some commission pixels with relatively low spectral values in
the SWIR band and low temperature in the thermal band. The thresholds
used in this approach were determined by measuring spectral deviation from
the mean value of the input images. These methods can accurately detect
cloud for the reported images, but the thresholds may not be transferable to
other images. To avoid confusion between bright surfaces and haze/cloud,
Chen et al. (2015) proposed an Iterative Haze Optimized Transformation
(IHOT) for improving haze/clouds detection for Landsat images with the
help of a corresponding clear image. By integrating an iterative procedure
of regressions into the HOT (Zhang etal., 2002), the reectance difference
between hazy and clear images, and reectance of hazy and clear images,
the land surface information can be removed. The IHOT result is derived
to characterize the haze contamination on Landsat images spatially. These
proposed approaches are practical and straightforward only using two-date
or multi-date Landsat images but heavily dependent on the quality and
availability of reference images. Besides, these approaches may not work well
if extensive land cover changes occurred between the acquisition dates of the
reference and cloudy images.
With free and open access to the Landsat archive, time series analysis
with Landsat images became possible, providing a new way to detect clouds
based on higher frequency Landsat observations. The LTS itself can be
used for detecting clouds. Goodwin etal. (2013) used LTS from TM/ETM+
to detect clouds. By using the minimum and median values of the blue
band as a reference, this algorithm can produce better cloud masks across
Queensland compared to Fmask (Zhu and Woodcock, 2012). However, it
has not yet been tested in environments with different soils, vegetation
16 Remote Sensing Time Series Image Processing
cover, and structure or areas with snow/ice cover (Goodwin et al., 2013).
Specically designed for monitoring land cover change, an algorithm called
Tmask (multitemporal mask) has been developed for automated masking of
cloud and snow for LTS (Zhu and Woodcock, 2014). This method estimated
time series models for each pixel based on “clear-sky LTS” previously
ltered by the Fmask algorithm. By using a robust tting approach, the
cloud observations that are missed by Fmask will have minimal impacts on
the estimation of the time series models. By comparing the model estimates
with actual Landsat observations for the green, NIR, and SWIR bands,
we will be able to detect any remaining cloud observations for the entire
stack of Landsat images. In addition to the Landsat images, there are also
algorithms developed for Landsat-like data, such as VENµS and Sentinel-2.
Hagolle etal. (2010) developed the Multi-Temporal Cloud Detection (MTCD)
method that detects sudden increases of reectance in the blue band on a
pixel-by-pixel basis using time series observations and tested the linear
correlation of pixel neighborhoods taken from pairs of images acquired
successively. The MTCD method provides better discrimination of cloudy
and clear sky pixels than the ACCA method for Landsat images. However, it
requires satellite data with high revisit frequency and sequential processing
of the data.
1.4.2 Cloud Shadow Detection Based on Multitemporal Landsat Images
Most cloud shadow detection algorithms using multitemporal Landsat
images assume that the presence of cloud shadows will lead to darker, colder,
and smoother features than the regular land surface (Irish, 2000; Le Hégarat-
Mascle and André, 2009). Wang etal. (1999) presented a wavelet transform
approach to detect cloud shadows for two Landsat TM images automatically.
Considering that the brightness changes of the cloud shadow-obscured
regions are much smoother than the regions with no shadows, the absolute
wavelet coefcients corresponding to cloud shadows decrease much greater
amount than those of other regions. Thus, a relative contrast difference for
the added result of the wavelet transform outputs was directly used to detect
cloud shadow for the two Landsat images with a static threshold. Different
from this complicated approach, Jin etal. (2013) detected the cloud shadows
simply by differencing the SWIR and the thermal bands from two-date
Landsat images and employed the geometric relationship between clouds
and their corresponding shadows to reduce false positive errors. Zhu and
Woodcock (2014) also identied cloud shadows for LTS by image differencing.
The reference values were predicted using a time series model for each pixel.
Though there are only a few cloud shadow detection approaches using
multitemporal Landsat images, these methods can provide better results than
the approaches based on a single-date Landsat image, especially for shadows
from thin clouds (Zhu and Woodcock, 2014).
17Cloud and Cloud Shadow Detection for Landsat Images
1.5 Discussions
1.5.1 Comparison of Different Algorithms
With so many different cloud and cloud shadow detection algorithms
available in the literature, it is essential to compare those approaches
and to provide further guidance on the selection of algorithms for those
interested in using LTS. A list of most of the automated cloud and cloud
shadow detection algorithms can be found in Table 1.3. We observe that the
most widely used detection algorithms are based on a single-date Landsat
image, probably due to the ease of implementation. Recently, Foga etal.
(2017) compared the performances of several popular algorithms using
278 unique cloud validation masks over the entire globe and found that
the CFmask (Fmask algorithm programmed in C) has the best overall
accuracy for Landsat data. It should be noted that the methods based on
multitemporal Landsat images can provide more accurate detection of cloud
and cloud shadow, which is especially important for time series analysis
(e.g., forest disturbance, land cover change, etc.) (Goodwin etal., 2013; Zhu
and Woodcock, 2014).
1.5.2 Challenges
Clouds are easily confused with snow/ice, especially for mountaintop snow/
ice (Selkowitz and Forster, 2015). These kinds of commissions can be reduced
by the NDSI threshold (Zhu and Woodcock, 2012), verication of clouds with
their corresponding shadows (Choi and Bindschadler, 2004), temperature
normalization (Qiu etal., 2017), or composition of temporal pixels in summer
season (Selkowitz and Forster, 2015). However, it is still difcult to separate
clouds from snow in some circumstances (e.g., icy clouds).
The cloud shadow detection accuracy is still relatively low. The geometry
projection of cloud is a good way to detect cloud shadow, but relies heavily on
the previously identied cloud masks, which have commission or omission
errors and subsequently result in inaccurate cloud shadows. In addition, the
cloud shadows are commonly confused with other dark features, such as
wetlands, dark urban, and terrain shadows. Terrain shadows can be removed
using the topographic correction model with the aid of DEMs (Jin etal., 2013;
Braaten etal., 2015; Qiu etal., 2017). The misidentication of cloud shadow
contributed from other dark features can also be corrected based on the
contextual information from the clouds’ heights estimated from neighboring
clouds (Qiu etal., 2017).
The use of multitemporal Landsat images can produce better cloud and
cloud shadow masks by differencing new observations with reference
observations. However, this kind of approach may not work well due to the
range of non-cloud related variations in reectance, such as the illumination
18 Remote Sensing Time Series Image Processing
geometry, land surface change, geometric misregistration, and variation
in radiometry or atmospheric composition (Hagolle et al., 2010; Goodwin
etal., 2013; Zhu and Woodcock, 2014). Furthermore, these algorithms are
computationally expensive compared to cloud and cloud shadow detection
algorithms based on a single-date Landsat image.
1.5.3 Future Development
1.5. 3.1 Spatial Information
When designing cloud and cloud shadow detection algorithms for Landsat
images, the spectral information and the temporal information have been
explored extensively, but the information contained in the spatial domain is
less studied (Gurney, 1982; Martins etal., 2002). We expect that more cloud and
cloud shadow detection algorithms will focus on the spatial characteristics of
clouds and their shadows and provide masks at higher accuracies.
1.5.3.2 Temporal Frequency
The approaches based on multitemporal Landsat images can provide more
accurate cloud and cloud shadow masks, when compared to the single-date
approaches (Goodwin et al., 2013; Zhu and Woodcock, 2014). In addition to
Landsat data, other Landsat-like satellites have also been launched, such
as Sentinel-2A/2B. The integration of multi-source images will allow more
frequent observations and further improve the detection accuracy. One major
restriction of the multitemporal cloud and cloud shadow detection algorithms
is that these algorithms require large amounts of data and computation time.
However, this will be less an issue with the rapid development of computation
tec h nolog y.
1.5.3.3 Haze/Thin Cloud Removal
Compared with thick clouds, thin clouds are transparent, and images
covered by thin clouds include information from both the atmospheric and
the ground underneath (Li et al., 2012). This gives us the opportunity to
remove the impacts of haze/thin clouds. If the satellite sensor prole and
the atmospheric properties are known, haze/thin clouds’ impacts can be
reduced by atmospheric correction (Vermote and Saleous, 2007). However, it
is difcult to acquire all the atmospheric properties (Liang etal., 2001), and
atmospheric correction may fail in handling the locally concentrated thin
clouds (Shen etal., 2014). Methods based on multispectral transformation,
such as Tasseled Cap (TC) transformation (Richter, 1996), HOT (Zhang
etal., 2002), and Advanced HOT (AHOT) (Liu etal., 2011), can remove haze/
thin clouds’impacts effectively. Besides, haze/thin clouds are generally
distributed in the low frequency parts of the image, which can be removed
by using alow-pass lter (Shen etal., 2014), such as Wavelet Analysis (WA)
19Cloud and Cloud Shadow Detection for Landsat Images
(Du et al., 2002) and Homomorphic Filter (HF) (Fan and Zhang, 2011).
While many haze/thin cloud removal methods are available, there are still
difculties in automated identication of haze/thin clouds using current
cloud detection algorithms. This will hamper the broad applications of haze/
thin cloud removal approaches.
1.6 Conclusion
Clouds and cloud shadows are a pervasive, dynamic, and unavoidable issue
in Landsat images, and their accurate detection is the fundamental basis
for analyzing LTS. Many cloud and/or cloud shadow detection algorithms
have been proposed in the literature. For cloud detection, most approaches
are based on a single-date Landsat image, which rely on physical-rules or
machine-learning techniques. With the policy of free and open Landsat
data, some automated cloud detection methods were developed based on
multitemporal Landsat images and can achieve better results. For cloud
shadow detection, the geometry-based approach is widely used in the single-
date algorithms. Meanwhile, by using multitemporal Landsat images, some
researchers used the image differencing method to better identify cloud
shadow. In this chapter, we reviewed many automated cloud and cloud
shadow detection algorithms, which can provide guidance on the selection
of algorithms for those interested in using LTS.
References
Ackerman, S. A., Strabala, K.I., Menzel, W. P., Frey, R. A., Moeller, C.C., and Gumley, L.E.
1998. Discriminating clear sky from clouds with MODIS. Journal of Geophysical
Research 103(D24): 32141–32157. doi:10.1029/1998J D200032 .
Berendes, T., Sengupta, S. K., Welch, R. M., Wielicki, B. A., and Navar, M. 1992.
Cumulus cloud base height estimation from high spatial resolution Landsat
data: A Hough transform approach. IEEE Transactions on Geoscience and Remote
Sensing 30(3): 430 – 443. doi:10.1109/36.142921.
Braaten, J. D., Cohen, W. B., and Yang, Z. 2015. Automated cloud and cloud shadow
identication in Landsat MSS imagery for temperate ecosystems. Remote Sensing
of Environment 169: 128–138. doi:10.1016/j.rse.2015.08.006.
Chen, J. M. and Cihlar, J. 1996. Retrieving leaf area index of boreal conifer forests
using Landsat TM images. Remote Sensing of Environment 55(2): 153–162.
doi:10.1016/003 4-4257(95)00195- 6.
Chen, S., Chen, X., Chen, J., and Jia, P. 2015. An iterative haze optimized transformation
for automatic cloud/haze detection of Landsat imagery. IEEE Transactions on
Geoscience and Remote Sensing 54(5): 2682–2694. doi:10.1109/TGRS.2015.2504369.
20 Remote Sensing Time Series Image Processing
Choi, H. and Bindschadler, R. 2004. Cloud detection in Landsat imagery of ice sheets
using shadow matching technique and automatic normalized difference snow
index threshold value decision. Remote Sensing of Environment 91(2): 237–242.
doi:10.1016/j.rse.2004.03.007.
Collins, J. B. and Woodcock, C. E. 1996. An assessment of several li near change detect ion
techniques for mapping forest mortality using multitemporal Landsat TM data.
Remote Sensing of Environment 56(1): 66–77. doi:10.1016/0034-4257(95)00233-2.
Derrien, M., Farki, B., Harang, L., LeGleau, H., Noyalet, A., Pochic, D., and Sairouni, A.
1993. Automatic cloud detection applied to NOAA-11 /AVHRR imagery. Remote
Sensing of Environment 46(3): 246–267. doi:10.1016/0034-4257(93)90046-Z.
Du, Y., Guindon, B., and Cihlar, J. 2002. Haze detection and removal in high resolution
satellite image with wavelet analysis. IEEE Transactions on Geoscience and Remote
Sensing 40(1): 210–217. doi:10.1109/36.981363.
Fan, C. N. and Zhang, F. Y. 2011. Homomorphic ltering based illumination
normalization method for face recognition. Pattern Recognition Letters 32(10):
1468–1479. doi:10.1016/j.patrec.2011.03.023.
Fassnacht, K. S., Gower, S. T., MacKenzie, M. D., Nordheim, E. V., and Lillesand, T. M.
1997. Estimating the leaf area index of North Central Wisconsin forests using
the Landsat thematic mapper. Remote Sensing of Environment 61(2): 2 29–24 5.
doi:10.1016/S0034-4257(97)00005-9.
Foga, S., Scaramuzza, P. L., Guo, S., Zhu, Z., Dilley, R. D., Beckmann, T., Schmidt, G.,
Dwyer, J., Hughes, M., and Laue, B. 2017. Cloud detection algorithm comparison
and validation for operational Landsat data products. Remote Sensing of
Environment 194: 379–390. doi:10.1016/j.rse.2017.03.026.
Goodwin, N. R., Collett, L. J., Denham, R. J., Flood, N., and Tindall, D. 2013. Cloud and
cloud shadow screening across Queensland, Australia: An automated method
for Landsat TM/ETM+ time series. Remote Sensing of Environment 134: 50–65.
doi:10.1016/j.rse.2013.02.019.
Gurney, C. M. 1982. The use of contextual information to detect cumulus clouds and
cloud shadows in Landsat data. International Journal of Remote Sensing 3(1): 51– 62.
doi:10.1080/01431168208948379.
Hagolle, O., Huc, M., Pascual, D. V., and Dedieu, G. 2010. A multi-temporal method for
cloud detection, applied to FORMOSAT-2, VENµS, LANDSAT and SENTINEL-2
images. Remote Sensing of Environment 114(8): 1747–1755. doi:10.1016/j.
rse.2010.03.002.
Homer, C., Huang, C., Yang, L., Wylie, B., and Coan, M. 2004. Development of a 2001
national land-cover database for the United States. Photogrammetric Engineering
and Remote Sensing 70(7): 829–840. doi:10.14358/PERS.70.7.829.
Huang, C. et al. 2010. Automated masking of cloud and cloud shadow for forest
change analysis using Landsat images. International Journal of Remote Sensing
31(20): 5449–5464. doi:10.1080/01431160903369642.
Hughes, M. J. and Hayes, D. J. 2014. Automated detection of cloud and cloud shadow in
single-date Landsat imagery using neural networks and spatial post-processing.
Remote Sensing of Environment 6(6): 4907–4926. doi:10.3390/rs6064907.
Irish, R. R. 2000. Landsat 7 automatic cloud cover assessment. Proceedings Volume 4049,
Algorithms for Multispectral, Hyperspectral, and Ultraspectral Imagery, Orlando,
FL, United States, AeroSense: International Society for Optics and Photonics.
doi:10.1117/12.410358.
21Cloud and Cloud Shadow Detection for Landsat Images
Irish, R. R., Barker, J. L., Goward, S. N., and Arvidson, T. 2006. Characterization of
the Landsat-7 ETM+ Automated Cloud-Cover Assessment (ACCA) algorithm.
Photogrammetric Engineering and Remote Sensing 72(10): 1179 1188. doi :10.14358/
PER S.72.10.1179.
Jin, S., Homer, C., Yang, L., Xian, G., Fry, J., Danielson, P., and Townsend, P. A. 2013.
Automated cloud and shadow detection and lling using two-date Landsat
imagery in the USA. International Journal of Remote Sensing 34(5): 1540–1560.
doi :10.1080/01431161.2012.720045.
Lee, J., Weger, R. C., Sengupta, S. K., and Welch, R. M. 1990. A neural network approach
to cloud classication. IEEE Transactions on Geoscience and Remote Sensing 28(5):
846–855. doi:10.1109/36.58972.
Le Hégarat-Mascle, S. and André, C. 2009. Use of Markov random elds for automatic
cloud/shadow detection on high resolution optical images. ISPRS Journal of
Photogrammetry and Remote Sensing 64(4): 3 51–36 6. do i:10.1016/j.i spr sjpr s.20 08.12.00 7.
Li, Q., Lu, W., Yang, J., and Wang, J. Z. 2012. Thin cloud detection of all-sky images
using Markov random elds. IEEE Geoscience and Remote Sensing Letters 9(3):
417–421. doi:10.1109/LGRS.2011.2170953.
Liang, S., Fang, H., and Chen, M. 2001. Atmospheric correction of Landsat ETM+ land
surface imagery. I. Methods. IEEE Transactions on Geoscience and Remote Sensing
39(11): 2490–2498. doi:10.1109/36.964986.
Liu, C. B., Hu, J. B., Lin, Y., Wu, S. H., and Huang, W. 2011. Haze detection, perfection
and removal for high spatial resolution satellite imagery. International Journal of
Remote Sensing 32(23): 8685 8697. doi:10.1080/01431161.2010. 547884.
Loveland, T. R. and Dwyer, J. L. 2012. Landsat: Building a strong future. Remote
Sensing of Environment 122: 22–29. doi:10.1016/j.rse.2011.09.022.
Lu, D. 2005. Aboveground biomass estimation using Landsat TM data in the
Brazilian Amazon. International Journal of Remote Sensing 26(12): 2509–2525.
doi:10.1080/014311605 0 0142145.
Luo, Y., Trishchenko, A. P., and Khlopenkov, K. V. 2008. Developing clear-sky, cloud
and cloud shadow mask for producing clear-sky composites at 250-meter spatial
resolution for the seven MODIS land bands over Canada and North America.
Remote Sensing of Environment 112(12): 4167– 4185. doi:10.1016/j.rse.2008.06.010.
Martins, J. V., Tanré, D., Remer, L., Kaufman, Y., Mattoo, S., and Levy, R. 2002.
MODIS cloud screening for remote sensing of aerosols over oceans using
spatial variability. Geophysical Research Letters 29(12): MOD4-1–MOD4-4.
doi:10.1029/2001GL013252.
Oreopoulos, L., Wilson, M. J., and Várnai, T. 2011. Implementation on Landsat data of a
simple cloud-mask algorithm developed for MODIS land bands. IEEE Geoscience
and Remote Sensing Letters 8(4): 597–601. doi:10.1109/LGRS.2010.2095409.
Pugmacher, D., Cohen, W. B., and Kennedy, R. E. 2012. Using Landsat-derived
disturbance history (1972–2010) to predict current forest structure. Remote
Sensing of Environment 122: 146 –165. doi:10.1016/j.rse.2011.09.025.
Potapov, P., Turubanova, S., and Hansen, M. C. 2011. Regional-scale boreal forest
cover and change mapping using Landsat data composites for European Russia.
Remote Sensing of Environment 115(2): 5 48561. doi:10.1016/j.r se.2010.10.001.
Qiu, S., He, B., Zhu, Z., Liao Z., and Quan, X. 2017. Improving Fmask cloud and cloud
shadow detection in mountainous area for Landsats 4–8 images. Remote Sensing
of Environment 199: 107–119. doi:10.1016/j.rse.2017.07.002.
22 Remote Sensing Time Series Image Processing
Richter, R. 1996. A spatially adaptive fast atmospheric corre ction algorithm. International
Journal of Remote Sensing 17(6): 1201–1214. doi:10.1080/01431169608949077.
Roy, D. P., Ju, J., Kline, K., Scaramuzza, P. L., Kovalskyy, V., Hansen, M., Loveland, T. R.,
Vermote, E., and Zhang, C. 2010. Web-enabled Landsat Data (WELD): Landsat
ETM+ composited mosaics of the conterminous United States. Remote Sensing
of Environment 114(1): 35–49. doi:10.1016/j.rse.2009.08.011.
Scaramuzza, P. L., Bouchard, M. A., and Dwyer, J. L. 2012. Development of the Landsat
data continuity mission cloud-cover assessment algorithms. IEEE Transactions
on Geoscience and Remote Sensing 50(4): 1140–115 4. do i:10.1109/TGR S.2011.2164 087.
Selkowitz, D. J. and Forster, R. R. 2015. An automated approach for mapping persistent
ice and snow cover over high latitude regions. Remote Sensing 8(1): 16. doi:10.3390/
rs8010016.
Shen, H., Li, H., Qian, Y., Zhang, L., and Yuan, Q. 2014. An effective thin cloud removal
procedure for visible remote sensing images. ISPRS Journal of Photogrammetry
and Remote Sensing 96: 224–235. doi:10.1016/j.isprsjprs.2014.06.011.
Simpson, J. J., Jin, Z., and Stitt, J. R. 2000. Cloud shadow detection under arbitrary
viewing and illumination conditions. IEEE Transactions on Geoscience and Remote
Sensing 38(2): 972–976. doi:10.1109/36.841979.
Sun, L. et al. 2016. A universal dynamic threshold cloud detection algorithm
(UDTCDA) supported by a prior surface reectance database. Journal of
Geophysical Research: Atmospheres 121(12): 7172–7196. doi:10.1002/2015JD02472 2.
U.S. Geological Survey. 2016a. L7 Irish Cloud Validation Masks. U.S. Geological Survey
data release. doi:10.5066/F7XD0ZWC. (accessed September 26, 2017)
U.S. Geological Survey. 2016b. L8 SPARCS Cloud Validation Masks. U.S. Geological
Survey data release. doi:10.5066/F7FB5146. (accessed September 26, 2017)
U.S. Geological Survey. 2016c. L8 Biome Cloud Validation Masks. U.S. Geological
Survey data release. doi:10.5066/F7251GDH. (accessed September 26, 2017)
Vermote, E., Justice, C., Claverie, M., and Franch, B. 2016. Preliminary analysis of
the performance of the Landsat 8/OLI land surface reectance product. Remote
Sensing of Environment 185: 4 6 56. doi:10.1016/j.r se. 2016.04.008.
Vermote, E. and Saleous, N. 2007. LEDAPS Surface Reectance Product Description.
College Park: University of Maryland. doi:null
Wang, B., Ono, A., Muramatsu, K., and Fujiwara, N. 1999. Automated detection
and removal of clouds and their shadows from Landsat TM images. IEICE
Transactions on Information and Systems 82(2): 453–460. doi:null
Wilson, M. J. and Oreopoulos, L. 2013. Enhancing a simple MODIS CLOUD mask
algorithm for the Landsat data continuity mission. IEEE Transactions on
Geoscience and Remote Sensing 51(2): 723–731. doi:10.1109/TGRS.2012.2203823.
Woodcock, C. E. etal. 2008. Free access to Landsat imagery, Science 320: 1011–1012.
doi:10.1126/scienc e. 320.5879.1011a.
Wulder, M. A., Masek, J. G., Cohen, W. B., Loveland, T. R., and Woodcock, C. E. 2012.
Opening the archive: How free data has enabled the science and monitoring
promise of Landsat. Remote Sensing of Environment 122: 2–10. doi:10.1016/j.
rse.2012.01.010.
Wulder, M. A., White, J. C., Loveland, T. R., Woodcock, C. E., Belward, A. S., Cohen,
W.B., Fosnight, E. A., Shaw, J., Masek, J. G., and Roy, D. P. 2016. The global Landsat
archive: Status, consolidation, and direction. Remote Sensing of Environment 185:
271–28 3. doi:10.1016/j.rse.2015.11.032 .
23Cloud and Cloud Shadow Detection for Landsat Images
Xian, G., Homer, C., and Fry, J. 2009. Updating the 2001 National Land Cover
Database land cover classication to 2006 by using Landsat imagery change
detection methods. Remote Sensing of Environment 113(6): 1133 –1147. doi :10.1016/j.
rse.2009.02.004.
Yuan, F., Sawaya, K. E., Loeffelholz, B. C., and Bauer, M. E. 2005. Land cover
classication and change analysis of the Twin Cities (Minnesota) Metropolitan
Area by multitemporal Landsat remote sensing. Remote Sensing of Environment
98(2): 317–328. doi:10.1016/j.rse.2005.08.006.
Zhang, Y., Guindon, B., and Cihlar, J. 2002. An image transform to characterize and
compensate for spatial vari ations in t hin cloud contamination of Landsat images.
Remote Sensing of Environment 82(2): 173–187. doi:10.1016/S0034-4257(02)00034-2.
Zheng, D., Rademacher, J., Chen, J., Crow, T., Bresee, M., Le Moine, J., and Ryu, S. R.
2004. Estimating aboveground biomass using Landsat 7 ETM+ data across a
managed landscape in northern Wisconsin, USA. Remote Sensing of Environment
93(3): 402–411. doi:10.1016/j.rse.2004.08.008.
Zhou, G., Zhou, X., Yue, T., and Liu, Y. 2016. An optional threshold with SVM cloud
detection algorithm and DSP implementation. ISPRS-International Archives of
the Photogrammetry, Remote Sensing and Spatial Information Sciences 41: 771–777.
doi:10.5194/isprs-archives-XLI-B8-771-2016.
Zhu, Z. and Woodcock, C. E. 2012. Object-based cloud and cloud shadow detection
in Landsat imagery. Remote Sensing of Environment 118: 83–94. doi:10.1016/j.
rse.2011.10.028.
Zhu, Z. and Woodcock, C. E. 2014. Automated cloud, cloud shadow, and snow
detection in multitemporal Landsat data: An algorithm designed specically
for monitoring land cover change. Remote Sensing of Environment 152: 217–234.
doi:10.1016/j.rse.2014.06.012.
Zhu, Z., Wang, S., and Woodcock, C. E. 2015. Improvement and expansion of the
Fmask algorithm: Cloud, cloud shadow, and snow detection for Landsats 4–7, 8,
and Sentinel 2 images. Remote Sensing of Environment 159: 269–27 7. doi:10.1016/j.
rs e.2014.12 .014.
... Shadows of buildings affect the applications of aerial images, such as urban change detection and traffic monitoring (see Adeline et al., 2013, and references therein). The screening of clouds and their shadows is an important step in the preprocessing of satellite imager data of, for example, Landsat and MODIS (see Zhu et al., 2018;Wang et al., 2019). Shadows degrade the quality of the images, lowering the accuracy of their applications such as land cover classification and change detection (see e.g. ...
... Supervised machine learning techniques (neural networks and support vector machines) have been proposed for cloud shadow detection in satellite images also (see e.g. Hughes and Hayes, 2014;Ibrahim et al., 2021), but they are generally computationally expensive, require large training data sets with classified shadows (which itself is the problem to be solved), and trained classifiers may not work for new scenes with different shadow patterns (Adeline et al., 2013;Zhu et al., 2018). ...
... Goodwin et al. (2013), Zhu and Woodcock (2014), Candra et al. (2016), and Candra et al. (2019) chose to perform spectral tests based on the reflectance differences with cloud-free historical reference images, for Landsat cloud shadow detection. Such multi-temporal shadow detection approaches generally enhance the shadow detection performance (Zhu et al., 2018), but they require the availability of cloud-free seasonally dependent reference images which may be challenging for satellites with long revisit periods. ...
Article
Full-text available
Cloud shadows are observed by the TROPOMI satellite instrument as a result of its high spatial resolution compared to its predecessor instruments. These shadows contaminate TROPOMI's air quality measurements, because shadows are generally not taken into account in the models that are used for aerosol and trace gas retrievals. If the shadows are to be removed from the data, or if shadows are to be studied, an automatic detection of the shadow pixels is needed. We present the Detection AlgoRithm for CLOud Shadows (DARCLOS) for TROPOMI, which is the first cloud shadow detection algorithm for a spaceborne spectrometer. DARCLOS raises potential cloud shadow flags (PCSFs), actual cloud shadow flags (ACSFs), and spectral cloud shadow flags (SCSFs). The PCSFs indicate the TROPOMI ground pixels that are potentially affected by cloud shadows based on a geometric consideration with safety margins. The ACSFs are a refinement of the PCSFs using spectral reflectance information of the PCSF pixels and identify the TROPOMI ground pixels that are confidently affected by cloud shadows. Because we find indications of the wavelength dependence of cloud shadow extents in the UV, the SCSF is a wavelength-dependent alternative for the ACSF at the wavelengths of TROPOMI's air quality retrievals. We validate the PCSF and ACSF with true-colour images made by the VIIRS instrument on board Suomi NPP orbiting in close proximity to TROPOMI on board Sentinel-5P. We find that the cloud evolution during the overpass time difference between TROPOMI and VIIRS complicates this validation strategy, implicating that an alternative cloud shadow detection approach using co-located VIIRS observations could be problematic. We conclude that the PCSF can be used to exclude cloud shadow contamination from TROPOMI data, while the ACSF and SCSF can be used to select pixels for the scientific analysis of cloud shadow effects.
... Therefore, it is critical to accurately label cloud-covered areas before further processing and analysis of optical imagery. Manuscript Cloud detection or pixel-by-pixel cloud labeling for satellite imagery has received increasing attention over the last few decades [11], [12]. Cloud detection methods for the Landsat imagery are commonly classified into two categories: physicalrule-and machine-learning-based methods [11]. ...
... Manuscript Cloud detection or pixel-by-pixel cloud labeling for satellite imagery has received increasing attention over the last few decades [11], [12]. Cloud detection methods for the Landsat imagery are commonly classified into two categories: physicalrule-and machine-learning-based methods [11]. Physical-rulebased methods detect clouds by identifying their physical or empirical characteristics, such as "bright," "white," "cold," and "high" [13]- [19]. ...
... The results showed that the ERF sizes and segmentation accuracies could largely differ for different cloud distributions (see Figs. [10][11][12]. For scenes with different relative percentages of thick clouds, both the region and boundary accuracies are substantially lower in the scenes with more thin clouds because the spectra of thin clouds are more difficult to distinguish from the land surface spectra. ...
Article
Deep semantic segmentation networks perform better in cloud detection of satellite imagery than traditional methods due to their ability to extract high-level features over a large receptive field. However, a large receptive field often leads to loss of spatial details and blurring of boundaries. Therefore, it is crucial to understand the role of the receptive field on the segmentation results, which has rarely been investigated for cloud detection tasks. This study, for the first time, explored the relationship between the receptive field size and the performance of a cloud detection network. Six typical networks commonly used for cloud detection and nine modified UNet variants with different depths, dilated convolutions, and skip connections were evaluated based on the Landsat 8 Biome (L8 Biome) dataset. The theoretical receptive field (TRF) and the effective receptive field (ERF) were introduced to measure the receptive field sizes of different networks. The results revealed a negative correlation between the ERF size and cloud segmentation accuracies for different cloud distributions and a relatively weak negative correlation between the TRF size and segmentation accuracies. Furthermore, ERFs were considerably smaller than the corresponding TRFs for most networks, implying that large-scale contextual information was not learned after training. This result indicates the importance of using networks with a small receptive field for cloud detection of Landsat 8 OLI imagery. Moreover, as the boundary accuracies are significantly lower than the region accuracies, future efforts should be devoted to addressing inaccurate boundary localization rather than exploring the contextual information over a large receptive field.
... This study obtained the seasonal NDVI value as the maximum of all available NDVI in the same season instead of the average value. We used the maximum rather than the average value to avoid the potential issues of broken and thin clouds, which may not be adequately masked out and would slightly reduce the value of NDVI [80,81]. It turned out that the seasonal NDVI time series allowed us to not only get NDVI values almost without contamination from clouds and aerosols at each pixel in the study area but also more accurately capture canopy transitions in the coastal wetland ecosystems. ...
... of broken and thin clouds, which may not be adequately masked out and would slightly reduce the value of NDVI [80,81]. It turned out that the seasonal NDVI time series allowed us to not only get NDVI values almost without contamination from clouds and aerosols at each pixel in the study area but also more accurately capture canopy transitions in the coastal wetland ecosystems. ...
Article
Full-text available
Sea-level rise and climate change stresses pose increasing threats to coastal wetlands that are vital to wildlife habitats, carbon sequestration, water supply, and other ecosystem services with global significance. However, existing studies are limited in individual sites, and large-scale mapping of coastal wetland degradation patterns over a long period is rare. Our study developed a new framework to detect spatial and temporal patterns of coastal wetland degradation by analyzing fine-scale, long-term remotely sensed Normalized Difference Vegetation Index (NDVI) data. Then, this framework was tested to track the degradation of coastal wetlands at the Alligator River National Wildlife Refuge (ARNWR) in North Carolina, United States, during the period from 1995 to 2019. We identified six types of coastal wetland degradation in the study area. Most of the detected degradation was located within 2 km from the shoreline and occurred in the past five years. Further, we used a state-of-the-art coastal hydrologic model, PIHM-Wetland, to investigate key hydrologic processes/variables that control the coastal wetland degradation. The temporal and spatial distributions of simulated coastal flooding and saltwater intrusion confirmed the location and timing of wetland degradation detected by remote sensing. The combined method also quantified the possible critical thresholds of water tables for wetland degradation. The remote sensing–hydrologic model integrated scheme proposed in this study provides a new tool for detecting and understanding coastal wetland degradation mechanisms. Our study approach can also be extended to other coastal wetland regions to understand how climate change and sea-level rise impact wetland transformations.
... Even if all clear observations have remained for analysis, the time series may have different temporal resolutions at different places and at different times due to the overlap of adjacent swaths, the presence of cloud or snow/ice, and the data acquisition strategies Zhu et al., 2018). For example, Gypsy Moth infestation usually only lasts for one or two months, and if all available Landsat data are used, we can have around four clear observations (without using observations from the neighboring path) in two months for most places (assume cloud cover is 50%). ...
Article
Full-text available
The discipline of land change science has been evolving rapidly in the past decades. Remote sensing played a major role in one of the essential components of land change science, which includes observation, monitoring, and characterization of land change. In this paper, we proposed a new framework of the multifaceted view of land change through the lens of remote sensing and recommended five facets of land change including change location, time, target, metric, and agent. We also evaluated the impacts of spatial, spectral, temporal, angular, and data-integration domains of the remotely sensed data on observing, monitoring, and characterization of different facets of land change, as well as discussed some of the current land change products. We recommend clarifying the specific land change facet being studied in remote sensing of land change, reporting multiple or all facets of land change in remote sensing products, shifting the focus from land cover change to specific change metric and agent, integrating social science data and multi-sensor datasets for a deeper and fuller understanding of land change, and recognizing limitations and weaknesses of remote sensing in land change studies.
... Unlike rule-based methods and multitemporal methods, machine learning-based cloud detection algorithms require a lot of training samples (Baetens et al., 2019;Zhu et al., 2019;Pérez-Suay et al., 2018). (Lee et al., 1990) proposed a texture-based neural network with four layers for cloud detection in Landsat MSS images, achieving 96% accuracy for Cirrus clouds. ...
Article
Full-text available
Cloud detection is a crucial step in the optical satellite image processing pipeline for Earth observation. Clouds in optical remote sensing images seriously affect the visibility of the background and greatly reduce the usability of images for land applications. Traditional methods based on thresholding, multi-temporal or multi-spectral information are often specific to a particular satellite sensor. Convolutional Neural Networks for cloud detection often require labeled cloud masks for training that are very time-consuming and expensive to obtain. To overcome these challenges, this paper presents a hybrid cloud detection method based on the synergistic combination of generative adversarial networks (GAN) and a physics-based cloud distortion model (CDM). The proposed weakly-supervised GAN-CDM method (available online https://github.com/Neooolee/GANCDM) only requires patch-level labels for training, and can produce cloud masks at pixel-level in both training and testing stages. GAN-CDM is trained on a new globally distributed Landsat 8 dataset (WHUL8-CDb, available online doi:https://doi.org/10.5281/zenodo.6420027) including image blocks and corresponding block-level labels. Experimental results show that the proposed GAN-CDM method trained on Landsat 8 image blocks achieves much higher cloud detection accuracy than baseline deep learning-based methods, not only in Landsat 8 images (L8 Biome dataset, 90.20% versus 72.09%) but also in Sentinel-2 images (“S2 Cloud Mask Catalogue” dataset, 92.54% versus 77.00%). This suggests that the proposed method provides accurate cloud detection in Landsat images, has good transferability to Sentinel-2 images, and can quickly be adapted for different optical satellite sensors.
... Over the past decades, with the release of massive amounts of optical satellite data and the continuous introduction of new data sources, many studies have focused on CCS detection for images taken by different sensors, and many CCS detection methods have been developed. Although the CCS detection literature has been reviewed in previous studies, they mainly focus on the categorization of methodologies (Goodman and Henderson-Sellers, 1988;Zhu et al., 2019a), cloud measuring equipment (Tapakis and Charalambides, 2013), and forms of CCS detection results (Mahajan and Fataniya, 2020). To further conduct a systematic summary of the current achievements and challenges in this field, this study comprehensively reviews the CCS detection literature from the features, algorithms, and validation perspectives as shown in Fig. 1. ...
Article
Full-text available
The presence of clouds prevents optical satellite imaging systems from obtaining useful Earth observation information and negatively affects the processing and application of optical satellite images. Therefore, the detection of clouds and their accompanying shadows is an essential step in preprocessing optical satellite images and has emerged as a popular research topic in recent decades due to the interest in image time series analysis and remote sensing data mining. This review first analyzes the trends of the field, summarizes the progress and achievements in the cloud and cloud shadow detection methods in terms of features, algorithms, and validation of results, and then discusses existing problems, and provides our prospects at the end. We aim at identifying the emerging research trends and opportunities, while providing guidance for selecting the most suitable methods for coping with cloud contaminated problems faced by optical satellite images, an extremely important issue for remote sensing of cloudy and rainy areas. In the future, expected improvements in accuracy and generalizability, the combination of physical models and deep learning, as well as artificial intelligence and online big data processing platforms will be able to further promote processing efficiency and facilitate applications of image time series. In addition, this review collects the latest open-source tools and datasets for cloud and cloud shadow detection and launches an online project (Open Satellite Image Cloud Detection Resources, i.e., OpenSICDR) to share the latest research outputs (https://github.com/dr-lizhiwei/OpenSICDR).
... For optical remote sensing data such as Landsat and EO-1, clouds, cloud shadows, and snow/ice need to be screened during image processing (Zhu et al., 2018). As for Landsat 7 and 8, we used the QA band to screen out the noise (e.g., clouds, cloud shadows, and snow/ice) for each pixel. ...
Article
Full-text available
Since 2017, the orbit of Landsat 7 has drifted outside its nominal mission requirement toward an earlier acquisition time because of limited onboard fuel resources. This makes quantitative analyses from Landsat 7 data potentially unreliable for many scientific studies. To comprehensively understand the effect of ongoing (2018–2020) orbit drift on Landsat 7 data, we compared surface reflectance and Top-Of-Atmosphere (TOA) reflectance of growing season observations (July 1 ± 30 days) from Landsat 7 with orbit drift and Landsat 8 with nominal orbit using a total of 10,000 randomly selected Northern Hemisphere (0–75⁰ N) terrestrial pixels. To evaluate the future (2021–2023) effect of Landsat 7's orbit drift, we analyzed the historical Northern Hemisphere terrestrial growing season Earth Observing-1 (EO-1) TOA reflectance images, which shared a similar orbit drift as Landsat 7 but occurred much earlier. Results suggest that Landsat 7's orbit drift has already led to a general decrease in surface reflectance and TOA reflectance in 2019 and 2020, with a limited impact (overall reflectance changes less than 0.007). The influence of orbit drift is more substantial for the two shortwave infrared (SWIR) bands and the near infrared (NIR) band, but less for the three visible bands (i.e., Red, Green, and Blue). The Normalized Difference Vegetation Index (NDVI), derived from either surface reflectance or TOA reflectance, increased less than 0.003 in 2020. According to the historical EO-1 TOA reflectance data, we estimate that the effect of Landsat 7's orbit drift will be much more dramatic in the future (e.g., the NIR and SWIR bands will decrease more than 0.015 since July 1, 2021), and for different land cover types, the effects of orbit drift are also quite different. To reduce this influence, we examined the c-factor Bidirectional Reflectance Distribution Function (BRDF) normalization approach to correct the orbit drift impact for Landsat 7 surface reflectance data collected between 2019 and 2020. We found that the c-factor BRDF can reduce the data difference substantially, but how this approach works after Landsat 7's orbit drifts further still requires more investigation. Therefore, we recommended that Landsat 7 can preserve its science capability until 2020, but will be less reliable for remote sensing applications that need accurate absolute radiometric values after 2020. Correction methods such as c-factor BRDF could be a potential viable approach to maintain its science capability going forward.
... The pixel QA band quality conditions are expressed as a cloud confidence level. In this study, only pixels with a clear condition (no cloud), low cloud confidence (LCC) and low cloud shadow (LCS) were used in the cloud masking method [32,51,52]. ...
Article
Full-text available
Total suspended sediment (TSS) is a water quality parameter that is used to understand sediment transport, aquatic ecosystem health, and engineering problems. The majority of TSS in water bodies is due to natural and human factors such as brought by river runoff, coastal erosion, dredging activities, and waves. It is an important parameter that should be monitored periodically, particularly over the dynamic coastal region. This study aims to monitor spatiotemporal TSS concentration over Teluk Lipat, Malaysia. To date, there are two commonly used methods to monitor TSS concentration over wide water regions. Firstly, field sampling is known very expensive and time-consuming method. Secondly, the remote sensing technology that can monitor spatiotemporal TSS concentration freely. Although remote sensing technology could overcome these problems, universal empirical or semiempirical algorithms are still not available. Most of the developed algorithms are on a regional basis. To measure TSS concentration over the different regions, a new regional algorithm needs to develop. To do so, two field trip was conducted in the study area concurrent with the passing of Landsat 8. A total of 30 field samples were collected from 30 sampling points during the first field trip and 30 samples from 30 samplings from the second field trip. The samples were then analyzed using an established method to develop the TSS algorithm. The data obtained from the first field trip were then used to develop a regional TSS algorithm using the regression analysis technique. The developed algorithm was then validated by using data obtained from the second field trip. The results demonstrated that TSS in the study area is highly correlated with three Landsat 8 bands, namely green, near-infrared (NIR), and short-wavelength (SWIR) bands, with R2 = 0.79. The TSS map is constructed using the algorithm. Analyses of the image suggest that the highest TSSs are mainly observed along the coastal line and over the river mouth. It suggested that the main contributing factors over the study area are river runoff and wave splash.
Chapter
This chapter presents an overview of the main time series analysis methods for environment monitoring with earth observation, from classical methods to the deep learning (DL) methods. It summarizes main differences between bi-temporal change detection, annual time series and dense time series analyses, and also presents the three main types of annual time series methods for environment monitoring. The chapter focuses on dense time series methods using all available data, first presenting the main data preprocessing requirements, and provides an overview of the four main types of change detection methods based on dense time series analysis. These include: map classification, trajectory classification, statistical boundary and ensemble approaches. The chapter discusses three kinds of network architectures suited for the analysis of satellite image time series (SITS): recurrent neural networks, convolutional neural networks and hybrid models combining both. It proposes a prospective reflection upon possible convergence at crossroads between SITS analysis, video processing, computer vision and DL.
Article
Full-text available
近年来,深度学习算法得到了长足的发展,并开始应用于云检测。但是深度神经网络模型参数众多,依赖大量训练样本,因此理解其泛化性能对于深度学习在不同遥感影像的实际应用具有重要的参考价值。本文以深度语义分割算法DeepLabv3+为例,以一组广泛使用的云标记数据集“L8 Biome”为验证数据,探讨该算法用于云检测时在不同的地表景观、空间分辨率和光谱波段组合的遥感影像上的泛化性能。云标记数据集“L8 Biome”包含96景具有全球代表性的Landsat 8 OLI影像及相应的人工云掩膜,被广泛用于测试云和云阴影检测算法性能。首先,利用Landsat 8 OLI云标记数据集“L8 Biome”,构建不同类型景观、不同空间分辨率、不同波段组合的训练影像集和测试影像集;其次基于不同训练样本集和测试集,评估了DeepLabv3+算法在不同情况下的云检测精度,并与Fmask算法作对比分析。研究结果表明:(1)使用全混合景观类型的训练集训练出来的云检测网络在总体检测精度(92.81%)与稳定度(标准差12.08%)上都优于使用单一景观类型的训练集训练得到的云检测网络,也优于Fmask的总体精度(88.75%)与稳定度(标准差17.34%),说明在构建深度学习算法的训练集时,应该尽可能包含多类型的地表景观;(2)将全混合景观训练集中剔除一类景观的样本(冰/雪景观除外)构建的“混合-1”训练集与全混合景观训练集训练的DeepLabv3+网络的云检测精度也相差不大,说明现有训练样本集已具备较强的景观泛化能力;(3)基于30 m空间分辨率的全混合景观训练样本集训练得到的DeepLabv3+云检测网络在不同分辨率(30 m、60 m、120 m、240 m)的测试集上云检测精度差异不大,都取得较好的效果,说明DeepLabv3+能够泛化应用于不同空间分辨率的遥感影像,相反Fmask直接应用于低分辨率影像时精度明显下降;(4)DeepLabv3+能充分自适应不同波段的信息用于云检测,总体来说更多的光谱波段输入能够提高DeepLabv3+的云检测的精度和稳定度,其中短波红外波段对于DeepLabv3+区分冰/雪与云具有重要价值,而热红外波段对DeepLabv3+云检测网络的性能提升很微小。以上结果说明利用现有数据集“L8 Biome”训练的DeepLabv3+云检测网络能够适用于多种类型的遥感影像,并优于Fmask算法。
Article
Full-text available
This paper presents a method which combines the traditional threshold method and SVM method, to detect the cloud of Landsat-8 images. The proposed method is implemented using DSP for real-time cloud detection. The DSP platform connects with emulator and personal computer. The threshold method is firstly utilized to obtain a coarse cloud detection result, and then the SVM classifier is used to obtain high accuracy of cloud detection. More than 200 cloudy images from Lansat-8 were experimented to test the proposed method. Comparing the proposed method with SVM method, it is demonstrated that the cloud detection accuracy of each image using the proposed algorithm is higher than those of SVM algorithm. The results of the experiment demonstrate that the implementation of the proposed method on DSP can effectively realize the real-time cloud detection accurately.
Article
Full-text available
Conventional cloud detection methods are easily affected by mixed pixels, complex surface structures, and atmospheric factors, resulting in poor cloud detection results. To minimize these problems, a new Universal Dynamic Threshold Cloud Detection Algorithm (UDTCDA) supported by a priori surface reflectance database is proposed in this paper. A monthly surface reflectance database is constructed using long-time-sequenced MODerate resolution Imaging Spectroradiometer surface reflectance product (MOD09A1) to provide the surface reflectance of the underlying surfaces. The relationships between the apparent reflectance changes and the surface reflectance are simulated under different observation and atmospheric conditions with the 6S (Second Simulation of the Satellite Signal in the Solar Spectrum) model, and the dynamic threshold cloud detection models are developed. Two typical remote sensing data with important application significance and different sensor parameters, MODIS and Landsat 8, are selected for cloud detection experiments. The results were validated against the visual interpretation of clouds and Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observation cloud measurements. The results showed that the UDTCDA can obtain a high precision in cloud detection, correctly identifying cloudy pixels and clear-sky pixels at rates greater than 80% with error rate and missing rate of less than 20%. The UDTCDA cloud product overall shows less estimation uncertainty than the current MODIS cloud mask products. Moreover, the UDTCDA can effectively reduce the effects of atmospheric factors and mixed pixels and can be applied to different satellite sensors to realize long-term, large-scale cloud detection operations.
Article
Full-text available
We developed an automated approach for mapping persistent ice and snow cover (glaciers and perennial snowfields) from Landsat TM and ETM+ data across a variety of topography, glacier types, and climatic conditions at high latitudes (above ~65�N). Our approach exploits all available Landsat scenes acquired during the late summer (1 August–15 September) over a multi-year period and employs an automated cloud masking algorithm optimized for snow and ice covered mountainous environments. Pixels from individual Landsat scenes were classified as snow/ice covered or snow/ice free based on the Normalized Difference Snow Index (NDSI), and pixels consistently identified as snow/ice covered over a five-year period were classified as persistent ice and snow cover. The same NDSI and ratio of snow/ice-covered days to total days thresholds applied consistently across eight study regions resulted in persistent ice and snow cover maps that agreed closely in most areas with glacier area mapped for the Randolph Glacier Inventory (RGI), with a mean accuracy (agreement with the RGI) of 0.96, a mean precision (user’s accuracy of the snow/ice cover class) of 0.92, a mean recall (producer’s accuracy of the snow/ice cover class) of 0.86, and a mean F-score (a measure that considers both precision and recall) of 0.88. We also compared results from our approach to glacier area mapped from high spatial resolution imagery at four study regions and found similar results. Accuracy was lowest in regions with substantial areas of debris-covered glacier ice, suggesting that manual editing would still be required in these regions to achieve reasonable results. The similarity of our results to those from the RGI as well as glacier area mapped from high spatial resolution imagery suggests it should be possible to apply this approach across large regions to produce updated 30-m resolution maps of persistent ice and snow cover. In the short term, automated PISC maps can be used to rapidly identify areas where substantial changes in glacier area have occurred since the most recent conventional glacier inventories, highlighting areas where updated inventories are most urgently needed. From a longer term perspective, the automated production of PISC maps represents an important step toward fully automated glacier extent monitoring using Landsat or similar sensors.
Article
Clouds are a pervasive and unavoidable issue in satellite-borne optical imagery. Accurate, well-documented, and automated cloud detection algorithms are necessary to effectively leverage large collections of remotely sensed data. The Landsat project is uniquely suited for comparative validation of cloud assessment algorithms because the modular architecture of the Landsat ground system allows for quick evaluation of new code, and because Landsat has the most comprehensive manual truth masks of any current satellite data archive. Currently, the Landsat Level-1 Product Generation System (LPGS) uses separate algorithms for determining clouds, cirrus clouds, and snow and/or ice probability on a per-pixel basis. With more bands onboard the Landsat 8 Operational Land Imager (OLI)/Thermal Infrared Sensor (TIRS) satellite, and a greater number of cloud masking algorithms, the U.S. Geological Survey (USGS) is replacing the current cloud masking workflow with a more robust algorithm that is capable of working across multiple Landsat sensors with minimal modification. Because of the inherent error from stray light and intermittent data availability of TIRS, these algorithms need to operate both with and without thermal data. In this study, we created a workflow to evaluate cloud and cloud shadow masking algorithms using cloud validation masks manually derived from both Landsat 7 Enhanced Thematic Mapper Plus (ETM +) and Landsat 8 OLI/TIRS data. We created a new validation dataset consisting of 96 Landsat 8 scenes, representing different biomes and proportions of cloud cover. We evaluated algorithm performance by overall accuracy, omission error, and commission error for both cloud and cloud shadow. We found that CFMask, C code based on the Function of Mask (Fmask) algorithm, and its confidence bands have the best overall accuracy among the many algorithms tested using our validation data. The Artificial Thermal-Automated Cloud Cover Algorithm (AT-ACCA) is the most accurate nonthermal-based algorithm. We give preference to CFMask for operational cloud and cloud shadow detection, as it is derived from a priori knowledge of physical phenomena and is operable without geographic restriction, making it useful for current and future land imaging missions without having to be retrained in a machine-learning environment.
Article
The surface reflectance, i.e., satellite derived top of atmosphere (TOA) reflectance corrected for the temporally, spatially and spectrally varying scattering and absorbing effects of atmospheric gases and aerosols, is needed to monitor the land surface reliably. For this reason, the surface reflectance, and not TOA reflectance, is used to generate the greater majority of global land products, for example, from the Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) sensors. Even if atmospheric effects are minimized by sensor design, atmospheric effects are still challenging to correct. In particular, the strong impact of aerosols in the visible and near infrared spectral range can be difficult to correct, because they can be highly discrete in space and time (e.g., smoke plumes) and because of the complex scattering and absorbing properties of aerosols that vary spectrally and with aerosol size, shape, chemistry and density.
Article
Most previous haze/cloud detection methods for Landsat imagery, e.g., haze optimized transformation (HOT), cannot adequately suppress land surface information and, in particular, often overestimate haze thickness over bright surfaces. This paper proposes an iterative HOT (IHOT) for improving haze detection with the help of a corresponding clear image. With an iterative procedure of regressions among HOT, the reflectance difference at the top of atmosphere (TOA) between hazy and clear images, and TOA reflectances of hazy and clear images, the land surface information can be removed, and the iterative HOT (IHOT) result is derived to spatially characterize the haze contamination in the Landsat images. A group of Landsat images that were acquired in different landscapes and seasons were used to test IHOT. Visual comparisons indicate that IHOT performed better than previous haze detection methods for images that were acquired in diverse landscapes and also performed robustly for hazy images that were acquired at different seasons when using the same reference clear image. Additionally, two indirect quantitative validations were used to illustrate that IHOT can provide the best transformation for accurately determining haze information. Therefore, it is expected that the proposed IHOT method will be used for automatic cloud/haze detection for large numbers of Landsat images if data sets of clear Landsat imagery are available.