Cirrus clouds play an important role in climate as they tend to warm the Earth–atmosphere system. Nevertheless their physical properties remain one of the largest sources of uncertainty in atmospheric research. To better understand the physical processes of cirrus clouds and their climate impact, enhanced satellite observations are necessary. In this paper we present a new algorithm, CiPS (Cirrus Properties from SEVIRI), that detects cirrus clouds and retrieves the corresponding cloud top height, ice optical thickness and ice water path using the SEVIRI imager aboard the geostationary Meteosat Second Generation satellites. CiPS utilises a set of artificial neural networks trained with SEVIRI thermal observations, CALIOP backscatter products, the ECMWF surface temperature and auxiliary data.
CiPS detects 71 and 95
By retrieving CALIOP-like cirrus properties with the large spatial coverage and high temporal resolution of SEVIRI during both day and night, CiPS is a powerful tool for analysing the temporal evolution of cirrus clouds including their optical and physical properties. To demonstrate this, the life cycle of a thin cirrus cloud is analysed.
High-level clouds cover 27–37
To capture the temporal evolution throughout the cirrus life
cycle as well as the diurnal cycles of cirrus coverage and
properties like cloud top height (CTH), ice optical thickness
(IOT) and ice water path (IWP), it is essential to accurately and
consistently detect and monitor cirrus during both day and
night. To this end, imagers like SEVIRI
Cirrus clouds can be detected from space-borne imagers
The CTH is an important variable as it determines the outgoing
longwave radiation. It can be retrieved from passive satellite
imagers during both day and night using e.g. radiance ratioing
(also referred to as
The limited amount of vertical information and sensitivity to
thin cirrus clouds is a recurrent drawback of passive
imagers. The space-borne lidar CALIOP
As an attempt to combine the advantages from a polar orbiting
lidar and a geostationary imager,
In this paper we present CiPS (Cirrus Properties from SEVIRI),
a new algorithm for cirrus remote sensing with SEVIRI that
exploits the basic idea of COCS: retrieving cirrus properties
using ANNs trained with CALIOP and SEVIRI data. However, CiPS
clearly differs from COCS in the implementation of this idea and
the achieved performance. For a more accurate cirrus detection
and determination of CTH and IOT, CiPS utilises a different set
of input parameters including numerical weather model data and
information about nearby pixels. In addition, CiPS classifies
each pixel as either cirrus-free, transparent cirrus or opaque
cirrus by means of dedicated classification ANNs. As CALIOP gets
saturated for thicker clouds, the opacity information is an
important additional piece of information in order to better
characterise the cirrus and the reliability of the ANN results
that was absent in COCS. Furthermore, CiPS is trained to retrieve
the IWP, resulting in a total of three climate relevant cirrus
cloud properties that can be estimated during both day and night
for the full SEVIRI field of view every 15
The remainder of this paper is divided into five parts. In
Sect.
SEVIRI is a passive imager operating aboard the geostationary
MSG satellites operational since
2004. SEVIRI is positioned at 0
CALIOP was launched as the main instrument aboard the CALIPSO
(Cloud-Aerosol Lidar and Infrared Pathfinder Satellite
Observations) satellite in 2006. CALIPSO is flying in
a sun-synchronous orbit as part of the A-Train
An artificial neural network consists of a number of neurons that
exchange information with each other, in a similar manner as
biological nerve cells transmit information via synapses in the
human brain. By assigning each neuron-neuron connection a numeric
tunable weight, the ANN has the ability to learn patterns and
approximate functions. The goal of an ANN is to derive a vector
of unknown output variables given a vector of known input
data. This tool is applied in Sects.
Generic structure of a multilayer perceptron (MLP), a form of a feed-forward artificial neural network used in this study.
In this study an MLP, a feed-forward
artificial neural network, is used. An MLP consists of three major
units; (1) the input layer, (2) the output layer and (3) the
hidden layer(s). The input layer holds as many neurons as input
variables and the output layer as many neurons as desired output
variables. The hidden layer(s) hold an arbitrary number of
additional neurons distributed over an arbitrary number of hidden
layers. All connections between the neurons within the MLP are in
the forward direction (input layer
When the MLP is given a vector of input data it uses the connection weights and possible biases to estimate the vector of output data. Thus, it is crucial that the weights and bias neurons are assigned correct values.
The weights are tuned by training the MLP, which is done with
a teacher–trainer approach, more known as supervised training.
A commonly used training algorithm is the back-propagation
algorithm. The most essential steps in the back-propagation
algorithm are explained below, but for the curious reader the
algorithm as a whole is well explained in
Using back-propagation the network is fed with a set of training
examples where the vector of input variables as well as the
vector of expected output variables are known. From the training
input data the MLP estimates its own output data using the
current weights. From the vector of
To find the minimum total error between the estimated and expected
output vectors for a complex problem and tune the weights
accordingly, a large training dataset is required. Training an MLP
is an iterative process, where each training example is presented
to the ANN multiple times until a satisfying result has been
achieved. With common ANN terminology the training completes one
While in recent years very potent new learning methods that are
based on back-propagation were developed, stochastic gradient
descent is still the most used method due to its simplicity and
robustness
Contingency table for the cirrus detection from CALIOP and CiPS.
This section introduces the validation metrics used for the validation later on in this paper.
The probability of detection (POD) is used to measure how
efficiently CiPS detects cirrus clouds and is given by
The mean percentage error (MPE) and mean absolute percentage error
(MAPE) are used to measure the accuracy of the CTH, IOT and IWP
retrievals with respect to CALIOP. The MPE is given by
The COCS
algorithm retrieves CTH and IOT of cirrus clouds from SEVIRI
COCS is an MLP with 10 input neurons (7 brightness temperatures
and temperature differences, viewing zenith angle, land–sea mask
and latitude), 2 output neurons (IOT and CTH) and 600 neurons in
one single hidden layer. COCS was trained with 3 years of data
including SEVIRI observations from both MSG-1 and MSG-2. The
detection of cirrus clouds takes place indirectly in COCS: a pixel
is cirrus-covered if its IOT
The V2 CALIOP L2 cloud layer products contain no information on
data quality and the feature classification flag and feature
optical thickness among other variables were released as beta
products (early release). V2 CALIOP layer data used in
CiPS is the new algorithm, based on the heritage from COCS in the sense that it also utilises artificial neural networks primarily trained with SEVIRI and CALIOP data. Significant enhancements with regards to the ANN structure, training input and output data and training methodology have been implemented in order to improve on retrieval performance and computational speed. In addition to CTH and IOT, CiPS is also trained to retrieve cloud opacity information and IWP.
In contrast to COCS, which uses one single ANN to retrieve IOT and
CTH, CiPS utilises four ANNs, making it possible to customise the
input variables, training data and ANN structures individually for
each task to be solved.
The first ANN is a classification network trained to detect
cirrus clouds using a binary cirrus cloud flag (CCF). Due to the
continuous activation function used by the ANN
(Sect. The second ANN is used for the CTH retrieval. The third ANN is used for the IOT/IWP retrieval. These two
variables are provided by the same network since they are physically
closely related CALIOP cannot provide accurate IOT/IWP retrievals for thicker
cirrus clouds where the laser beam is completely attenuated. Hence
the estimated IOT and IWP by CiPS for such situations should not be
trusted. Therefore a second classification network is introduced
with CiPS, trained to identify the cirrus clouds where CALIOP is
saturated. Similarly to the cirrus detection ANN, the opacity
classification ANN retrieves real numbers in the interval (0,1), which can be
regarded as an opacity probability information. From here a binary
opacity flag (OPF) is obtained using a suitable opacity
classification threshold (Sect.
The following subsections introduce all input data used to train
CiPS. An overview is provided in Table
Brightness temperatures from all thermal channels of SEVIRI except
for the ozone channel at 9.7
Input data used to train the four ANNs contained in CiPS. BT is brightness temperature, regavg is regional average, regmax is regional maximum and VZA is the viewing zenith angle.
For the classification ANNs (CCF and OPF) the regional
With CiPS we introduce modelled data from the ECMWF ERA-Interim
re-analysis dataset
Along with the data provided by SEVIRI and ECMWF, additional
auxiliary datasets are used. The latitude provides valuable
information about the geographical location with respect to the
global circulation convergence and divergence zones (e.g. the
ITCZ, subsidence regions and the polar front) which strongly
affect the presence and properties of cirrus clouds. Considering
the SEVIRI viewing zenith angle, the SEVIRI pixel size and slant
path length are implicitly accounted for. Two flags indicating the
presence of surface water and permanent ice/snow are
included to gain additional information about the observed surface
type. Due to the seasonal variations in the global circulation and
the presence of cirrus clouds
The cirrus presence and properties, including a CCF and an OPF as
well as the CTH, IOT and IWP, are derived from the V3 CALIOP L2
5
Even though the cloud and aerosol layer products are reported with
a spatial resolution of 5
The minimum detectable backscatter of CALIOP depends on the
scattering target (the cirrus cloud in this case), the altitude as
well as the vertical and horizontal averaging of the data
The improved quality of the V3 CALIOP products allows us to omit
the filtering processes applied to the V2 data used for COCS (see
Sect.
In the following, all quantities referring to CALIOP will be
denoted as IOT
The CALIOP products are chosen as training reference data for CiPS as they should provide the most accurate estimates of especially CTH but also IOT for thin cirrus clouds from space. It is important to note that an ANN can never be better than its training reference and all deficiencies and/or biases in the training reference data will be inherited by the ANN. Since possibly inherited artefacts of the ANN will not show when validated against independent CALIOP retrievals, one must be aware of the accuracy and limitations of the training data.
For transparent cirrus layers the agreement in IOT between CALIOP
and CPL is good with on average 15
The accuracy of the CALIOP IWC/IWP is directly related to the
accuracy of the extinction retrievals as well as the IWC
parameterisation from
To learn the relationship between the SEVIRI, ECMWF, auxiliary
data and the cirrus properties from CALIOP, an extensive dataset
is created containing spatial and temporal collocations of all
variables. The training dataset covers the time period from April
2007 to January 2013, which is the time when MSG-2 was the
operational satellite at 0.0
For this time period all quality-controlled CALIOP data within the
SEVIRI field of view are identified and collocated with single
SEVIRI pixels in time and space. Due to the different viewing
geometries of SEVIRI and CALIOP, the same cloud seen by SEVIRI and
CALIOP at the same time appears to be located at two different
positions. The magnitude of this displacement depends on the
viewing angle and the altitude of the cloud layer. This effect has
been corrected for using the latitude, longitude and cloud top
altitude from CALIOP (parallax correction) to project ice clouds
to the SEVIRI grid. The cirrus properties from CALIOP are
spatially collocated with SEVIRI observations from the pixel
having the largest overlap with the 5
When collocating SEVIRI and CALIOP observations with the purpose
of training an ANN one must consider two aspects. (1) Even though
the 5
The ECMWF surface temperatures are spatially collocated with the
satellite observations using nearest neighbour. For the temporal
collocation, the ECMWF re-analysis data are linearly interpolated
between the ECMWF 6
The relative number distribution of the cirrus IOT (
The full collocated dataset, covering the entire SEVIRI disc and
a time period of almost 6 years, contains close to 50 million
collocations. Of those collocations, 80
The remaining 20
To train and apply CiPS the Fast Artificial Neural Network library
Three hidden layers are used for the cirrus cloud detection, two
for the CTH and IOT and IWP retrievals and a single hidden layer
for the opacity classification. All ANNs use 16 hidden neurons per
hidden layer (see Sect.
The ANNs are initially trained using 25
To avoid overfitting, the error against the independent internal
validation datasets (Sect.
For each task/ANN the training is repeated twice in order to
reduce the risk of having a bad end performance as a result of
a bad set of initial weights (from Widrow and Nguyen's algorithm;
Using a common standard desktop PC (using 1 core
@ 3.40
The POD and FAR of the CiPS cirrus cloud detection and opacity classification ANNs as a function of classification threshold. The red circles indicate the final thresholds selected for the two ANNs.
The difference in accuracy between each MLP structure and the least complex MLP structure having one hidden layer with 16 hidden neurons (1–16).
As described in Sect.
When developing CiPS, several ANNs with different MLP structures
were trained in order to investigate the effect of the MLP
structure on the end performance and to determine the respective
structures that offer the best trade-off between accuracy and
application time. For each ANN contained in CiPS several networks
with different structures were trained using one, two and three
hidden layers with either 16 or 64 hidden neurons per hidden
layer. For the single hidden layer structures we also train with
128 hidden neurons. Also here the training was repeated twice for
each network in order to reduce the risk of having a bad end
performance as a result of a bad set of initial weights. Again,
only the best performing network among the two is further
evaluated after the training. All different structures were
trained according to the first phase as explained above
(Sect.
Figure
Approximate time required to process 1 million data points using the different ANN structures investigated in this study. The number to the left of the hyphen is the number of hidden layers and the number to the right the number of hidden neurons per hidden layer.
Furthermore, Table
In all cases, already small networks produce reasonable
results. In many cases differences between structures are not
very large. Nevertheless, we also see that larger ANNs can always
solve the problems in a more accurate way and especially for the
cirrus cloud detection it is beneficial to either use more hidden
neurons or add more hidden layers rather than using a simple
structure with one hidden layer and 16 hidden neurons
(1–16). Using two or three hidden layers with 64 hidden neurons
each (2–64, 3–64) yields a POD that is up to 8 percentage
points higher compared to one hidden layer with 16 hidden neurons
(1–16). Similarly, a structure with three hidden layers and 16
hidden neurons (3–16) yields a POD that is up to 5.5 percentage
points higher compared to the structure with one hidden layer and
16 hidden neurons (1–16). Although three hidden layers with 64
neurons each (3–64) offer the highest accuracy for all cases,
such a complex structure processes the data significantly slower
by a factor of 8 or 6 compared to the smaller structures with 2 or 3
hidden layers and 16 neurons per layer. For the IOT retrieval,
a larger ANN is mostly beneficial for the thinner cirrus and the
MAPE with respect to CALIOP seems to be saturated and hardly
improvable for IOT
In this section CiPS is applied to the 1 June 2015
12:30
Figure
Figure
In this section the performance of CiPS is validated against V3
CALIOP products using the 10
An in-depth characterisation of CiPS with respect to (1) the relative
importance of the different input variables, (2) the effect of the underlying
surface type as well as underlying liquid water clouds and aerosol layers on
the cirrus cloud retrieval, (3) the retrieval errors as a function of IOT and
CTH combined and (4) the sensitivity to radiometric noise in the SEVIRI input
data is presented in
The CCF of CiPS and COCS and the OPF of CiPS are evaluated as
a function of the geographic position. This aspect is interesting
due to the very different meteorological conditions present on the
SEVIRI disc. Figure
As mentioned in Sect.
COCS has an equally low FAR over arid regions but has a clearly
higher FAR in general. In particular over icy surfaces like
Greenland and Antarctica, COCS overestimates the cirrus presence,
with FARs up to approx. 90
Top: the FAR of the CCF retrieved by CiPS
The POD of CiPS and COCS as a function of the IOT retrieved by CALIOP.
Due to the high probability of cirrus cloud presence along the
ITCZ, the effect of the higher FAR of CiPS over this region is
small, since a high cirrus probability prevents false alarms from
occurring. Figure
The FAR can easily be optimised by reducing the number of detected
cirrus clouds (see Fig.
Figure
FAR of the CiPS OPF (opacity flag).
Density scatter plots with the CTH retrieved by
Figure
With CiPS the CTH is retrieved with a higher accuracy compared to COCS, especially for high and low cirrus clouds. The correlation between CALIOP and CiPS is 0.90. For CALIOP and COCS, the correlation coefficient is 0.82.
The MPE shows that CiPS overestimates and underestimates the CTH
of the lowest and highest cirrus clouds, respectively, even if the
errors are smaller than for COCS. From 8 to 15
Density scatter plots with the IOT retrieved by
The CTH has a strong latitude dependency and the CiPS results
shown in Fig.
The MPE shows a clear latitude dependency and in contrast to
Fig.
Note the difference between the CiPS CTH retrieval and standard
ones
Figure
For a better visualisation of the lower IOT range, where most points are
located, the density scatter plots have logarithmic axes. This does, however, visually reduce the errors, so for a quantitative evaluation
attention should be paid to Fig.
Figure
The scatter between
In contrast to the CTH
As expected and as seen in Figs.
As discussed in Sect.
In this section the potential of CiPS is illustrated by analysing
the temporal evolution of a thin cirrus cloud throughout its life
cycle. The life cycle of natural cirrus and contrails is an
important aspect to study
Here we analyse the life cycle of an outflowing cirrus
originating from an orographic cirrus. The cirrus cloud was
identified south of the Pyrenees on 26 September 2014 at
10:00
The path and temporal evolution of the cirrus cloud with
a temporal resolution of 120
Temporal evolution of the cloud properties for the cirrus
described in Fig.
The temporal evolution of the cloud horizontal area can be seen
at full temporal resolution (5
The cirrus cloud detaches from the orographic cirrus at
05:25
The IOT and IWP start to decrease at around 11:30
The CiPS algorithm presented in this paper detects cirrus clouds
and retrieves their CTH, IOT and IWP along with an OPF using SEVIRI, ECMWF and auxiliary data. CiPS utilises a set
of four artificial neural networks, trained with V3 CALIOP L2
layer data as a reference. CiPS does not take advantage of the
SEVIRI channels with significant solar contribution and can thus
be used during both day and night. By using ANNs, the idea is to
combine the high sensitivity and vertical resolution of CALIOP
with the large spatial coverage and high temporal resolution of
SEVIRI. Thus, the ultimate goal of CiPS is to retrieve CALIOP-like
cirrus properties for the full SEVIRI disc (approx. one-third of
the Earth) every 15
CiPS shows a good performance when validated against independent
CALIOP data. CiPS detects 95
CiPS has a better performance in all aspects with respect to COCS,
another algorithm that uses ANNs for retrieving the CTH and IOT
from SEVIRI using CALIOP as reference
As an application example, the life cycle of a thin cirrus cloud
and the temporal evolution of its properties is investigated. The
cirrus cloud lives for nearly 20
The approach of using ANNs is very fast and requires little
computational power compared to standard physical methods that
require extensive radiative transfer calculations and/or
interpolation in a multidimensional space. On a common standard
PC, one complete SEVIRI image with
With CiPS we are now able to study the temporal evolution, life cycles and diurnal cycles of thin cirrus clouds, natural and anthropogenic (contrails), including their coverage, CTH, IOT and IWP with a higher degree of accuracy. The inclusion of a physical variable like the IWP further allows for direct comparison with weather, climate or large eddy simulation models.
As a next step, the CiPS retrievals will be further characterised, for
example
with respect to the underlying surface type and the presence of
aerosol layers and liquid water clouds below the cirrus (see
MSG/SEVIRI L1.5 data are available at
The authors declare that they have no conflict of interest.
This research was supported by the DLR (Deutsches Zentrum für Luft- und Raumfahrt)/DAAD (Deutscher Akademischer Austauschdienst) Research Fellowship Programme für Doktoranden, 14.
We thank the NASA Atmospheric Science Data Center for their kind support and for providing the V3 CALIOP layer data in a subsetted form. We also thank Mark Vaughan for his guidance on how to properly account for the vertical overlap of cloud and aerosol features in the CALIOP layer products. We want to express our gratitude to Diego Loyola for an interesting and helpful discussion about the application of ANNs in satellite remote sensing. We also thank Stephan Kox for the discussion on COCS and the relevant routines that were provided. We gratefully acknowledge the constructive comments of three anonymous reviewers, Florian Ewald, André Butz and Ulrich Schumann that greatly improved the quality and clarity of this paper.
The SEVIRI data were provided by EUMETSAT (European Organisation for the Exploitation of Meteorological Satellites) and the modelled surface temperature was obtained from ECMWF (European Centre For Medium-Range Weather Forecasts). The MODIS MCD12C1 data product used to derive the land surface type flags was retrieved from the online Data Pool, courtesy of the NASA Land Processes Distributed Active Archive Center (LP DAAC), USGS/Earth Resources Observation and Science (EROS) Center, Sioux Falls, South Dakota. The article processing charges for this open-access publication were covered by a Research Centre of the Helmholtz Association. Edited by: Alexander Kokhanovsky Reviewed by: three anonymous referees