This paper presents the simultaneous retrieval of aerosol optical thickness
and surface properties from the CISAR algorithm applied both to geostationary
and polar-orbiting satellite observations. The theoretical concepts of the
CISAR algorithm have been described in

The retrieval of aerosol properties over land surfaces from space
observations is a challenging problem due to the strong radiative coupling
between atmospheric and surface radiative processes. Different approaches are
usually exploited to retrieve different Earth system components (e.g.

This paper is organised as follows. Section

The fundamental principle of OE is to maximise the
probability

Locations of selected AERONET stations. All stations are located within the SEVIRI field of view.

In order to evaluate the CISAR algorithm performance when applied to
observations acquired from different orbits, 20 AERONET stations located
within the SEVIRI field of view have been selected (Fig.

AERONET targets.

For each of these stations, satellite data have been acquired, together with
ancillary information, such as the cloud mask and the model parameters, which
are the parameters that are not retrieved by the algorithm but that influence
the observation. Satellite data and ancillary information are accumulated in
time to form a multi-angular observation vector

The observation uncertainty

MSG nominal position is 0

The PROBA-V satellite mission is intended to ensure the continuation of the
Satellite Pour l'Observation de la Terre 5 (SPOT5) VEGETATION products
starting from May 2014

SEVIRI (in blue), PROBA-V (in red) and MODIS (in green) spectral responses.

The similarities between the three SEVIRI solar bands and the red, NIR and
SWIR PROBA-V bands permit the evaluation and comparison of the CISAR
performances when applied to the two instruments, the spectral responses of
which are shown in Fig.

PROBA-V radiometric noise has been delivered by VITO (Sindy Sterckx, personal
communication, September 2017) per camera and per band (Table

PROBA-V instrument noise (%).

Total radiometric uncertainty median values (%).

The uncertainty

The total relative radiometric uncertainty median values are shown in Table

In addition to satellite observations, a cloud mask and the model parameters
are required. For SEVIRI observations, the nowcasting satellite application
facility (SAF) cloud mask

The uncertainties of the model parameters

Water vapour transmittance in the SEVIRI, PROBA-V and MODIS bands.

The median values of the equivalent model parameter noise (EQMPN), computed
as in Eq. (

Total EQMPN median values (%).

Within an OE framework, the definition of the prior information and its uncertainty plays a fundamental role. In CISAR four different sources of prior information are considered:

The corresponding prior uncertainty is defined as half of the variability
range of the solution

When

Polar plot of the angular sampling during 5 days
(1–5 May 2015) of SEVIRI observations (

FASTRE, the CISAR forward radiative transfer model (RTM), and its uncertainty

FASTRE relative uncertainty in the SEVIRI and PROBA-V processed bands (%).

In order to perform the inversion on actual satellite data, the observations
are accumulated in time and the corresponding uncertainty is computed as
described in Sect.

CISAR set-up parameters.

The definition of the first guess is an important aspect of the inversion
process and it is defined in order to minimise the possibility of finding
local minima. When a minimum value is found, an investigation of the cost
function in the vicinity of the solution should be made in order to determine
whether or not it is a local minimum. However, this exploration could be
computationally expensive. In order to minimise the possibility of local
minima without degrading the computational performances, the AOT first guess
is assigned to successive observations alternating between a low-value

Solution space (black triangle) for the wavelength 0.6

From the retrieved set of RPV parameters the BHR is calculated, assuming
perfectly diffuse illumination conditions, and the AOT is extrapolated at
0.55

The aerosol vertices subsample the entire solution space to a region where
the aerosol properties can be retrieved. The relationship between the
particle size and the single-scattering properties has been discussed in
Part 1. As recommended, three vertices are selected, defined by the asymmetry
factor

Histograms of the distribution of the Jacobians related to the RPV
parameters (

The analysis of the information content relies on a two-fold approach. First,
the Jacobians are used as an indicator of the TOA BRF sensitivity to state
variable changes under different observation conditions. Next, the entropy is
used as a rigorous metric to determine the information content of the
observation system for each radiometer. The Jacobians, i.e. the partial
derivatives of the forward model with respect to the state variables, are
affected by the changes in illumination and viewing geometry both in terms of
sign and magnitude

Distribution of the AOT-scaled Jacobian over Carpentras (dark surface) and Zinder Airport (bright surface). The histograms are obtained from PROBA-V observations (RED band) over the year 2015.

Median and standard deviation of the scaled Jacobians. The table
refers to all processed targets during 2015. The values are shown for the
SEVIRI and PROBA-V bands centred at 0.6

An illustrative example of the distributions of the Jacobians relative to the
RPV parameters is shown in Fig.

Scaled AOT Jacobians time series over Carpentras, France (vegetated
target) related to SEVIRI VIS0.6 band (

The aerosol contribution to the TOA BRF differs according to the brightness of the surface.
Figure

Table

Scaled AOT Jacobians (left

A more rigorous analysis of the information content can be made through the
entropy, which measures the uncertainty reduction

Distribution of the entropy related to the AOT (

In CISAR, the entropy is calculated considering the surface and atmospheric
state variables and their associated prior and posterior uncertainty
separately; the entropy distribution is shown in Fig.

Section

Correlation (in red) and RMSE (in blue) variations as a function of the mismatch between the satellite observation and the simulated signal (test 3). The figure refers to the CISAR AOT retrieval evaluation against AERONET data. These results are obtained from CISAR applied to SEVIRI observations.

A new approach is proposed for the CISAR algorithm, which combines a series
of individual tests

The first test to be performed is on the convergence
of the inversion. When the maximum number of iterations is reached,

The validity of the retrieved total AOT and of the
surface BHR is examined in tests 1 and 2. In CISAR, a validity range for
each state variable is defined, based on physical boundaries and empirical
observations. When the value of retrieved AOT (BHR) falls at the extremes of
this range,

As discussed in Sect.

Non-linear

Correlation (straight lines) and RMSE (dashed lines) variations as a function of the QI. The figure refers to the CISAR AOT retrieval from SEVIRI (in blue) and PROBA-V (in red) observations evaluated against AERONET data. The QI is rounded to the nearest 0.1.

CISAR-retrieved BHR from comparison of actual observations with MODIS in all the processed bands.

The magnitude of the Jacobians gives information on the sensitivity of the
TOA BRF to the state variables. Performing a test on the Jacobians related to
each state variable can be computationally expensive. In order to reduce the
computational effort, only the Jacobian of the AOT is taken into account. The
spectral constraints applied to the AOT variability as in Sect.

The aim of this test is to discard observations with little or no sensitivity
to the AOT and to identify those situations in which the test on the misfit is
successful because of the prior information and/or the temporal and spectral
constraints (Sect.

Section

CISAR-retrieved BHR from SEVIRI (blue dots) and PROBA-V (red dots)
and MODIS land product (green triangle) over Zinder Airport (Niger, Africa).
The results are shown for each sensor's band centred at 0.6

The final QI is computed by combining the results of the tests performed on
the retrieved solution:

The final QI

The CISAR BHR, computed from the RPV parameters, is compared with the MODIS
land product

Box plots showing the CISAR AOT retrieval extrapolated at
0.55

The CISAR AOT retrieval, extrapolated at 0.55

Figure

The box plots in Fig.

The overestimation of low AOT might originate from the different spatial
scales of the satellite observations and the ground measurements. Most of
the selected AERONET stations are located in Europe (Fig.

Fine- to coarse-mode ratio distribution at 0.6

The CISAR potential to discriminate between the fine and coarse mode is
analysed next. Figure

SSA distributions at 0.6

Same as
Figure

In Sect.

This paper describes and evaluates the application of the CISAR algorithm to satellite observations acquired from geostationary and polar-orbiting instruments. The theoretical aspects of CISAR, a new generic algorithm for the joint retrieval of surface reflectance and aerosol properties, with continuous variation of all the state variables in the solution space, are described in Part 1. In the latter, CISAR is applied to simulated noise-free observations in the principal plane. This paper provides an evaluation of the algorithm in non-ideal situations, i.e. actual satellite observations acquired from both geostationary and polar-orbiting satellites, namely SEVIRI and PROBA-V.

The proposed retrieval method relies on an OE approach which consists of the
inversion of FASTRE, a simple radiative transfer model composed of two
horizontal layers. The FASTRE model is evaluated in Sect.

The analysis of the information content of the satellite observations is
performed in Sect.

The CISAR retrieval is evaluated against independent data sets. The retrieved
AOT is compared to AERONET data. A specific QI has been developed to
disregard suspicious retrievals. With an RMSE of 0.162 for SEVIRI and 0.176
for PROBA-V, CISAR shows better performances when applied to the
geostationary satellite. CISAR retrieves the single-scattering aerosol
properties, assuming a linear behaviour of

Several aspects of the new CISAR algorithm would still require additional
effort to improve its performance. The analysis of the Jacobian median
values has revealed the very small magnitude of the fine- and coarse-mode AOT
Jacobians. The spectral and temporal constraints of the AOT variability play
a critical role in supporting the assessment of aerosol properties. However,
these constraints might lead to an underestimation of the AOT for large
values. The impact of cloud mask omission errors on the AOT overestimation at
low optical thickness deserves additional work. In order to reduce the impact
of cloud contamination in the AOT retrieval, a new version of the CISAR
algorithm is under development in the framework of the ESA-SEOM ConsIstent
Retrieval of Cloud Aerosol Surface (CIRCAS) project (

As pointed out in Part 1, the limited number of state variables retrieved by
CISAR allows the same algorithm to be applied to sensors which have not been
originally designed for aerosol or surface albedo retrieval. The possibility
to apply the same algorithm to data acquired by different instruments for the
retrieval of several essential climate variables (ECVs) presents a decisive advantage, as it provides
radiatively consistent ECVs derived from different sensors. Conversely, the
use of separate methods for the retrieval of different variables might lead
to a radiance bias, which has to be corrected before the assimilation of
these variables

Results presented in Sect. 6 are available from the authors upon request.

Included are the scatter plots of the BHR retrieved by CISAR versus the BHR
delivered by MODIS (Figs. S1, S2), and a few examples of the CISAR high-AOT
retrievals compared with AERONET data. The supplement related to this article is available online at:

ML adapted the CISAR algorithm to the two satellite observations, developed the QI in Sect. 5, validated the results and wrote the paper. YG wrote FASTRE and provided support for the scientific aspects of the inversion process and the application of CISAR to SEVIRI and PROBA-V observations.

The authors declare that they have no conflict of interest.

This work has been performed in the framework of ESA projects aerosol_cci2 and PV-LAC under the contracts 4000109874/14/I-NB and 4000114981/15/I-LG respectively. The authors would like to thank the reviewers for their fruitful suggestions. Edited by: Andrew Sayer Reviewed by: two anonymous referees