3.2 Cloud Property Remote Sensing with the APICS Algorithm
3.2.3 Cirrus Cloud Mask
The APICS retrieval is only performed for pixels classified as cirrus. This is mandatory since else a measurement of increased reflectivity compared to the pre-computed clear sky value at 0.6µm will be misinterpreted by the algorithm as a cirrus cloud, even if the considered pixel is completely cirrus free. Such increased reflectivity values can, for example, be caused by high aerosol loadings or water clouds.
APICS originally relied on a cirrus cloud mask generated by the “Meteosat Second Gener- ation Cirrus Detection Algorithm” v2 (MeCiDa) (Krebs et al., 2007; Ewald et al., 2013). However, it was foreseen that the retrieval of circumsolar radiation would benefit if the cirrus cloud mask used in conjunction with APICS was improved considering the detection efficiency of optically thin clouds. During the course of this study another cirrus cloud property retrieval algorithm was developed at DLR, which is called “Cloud Optical prop- erties derived from CALIOP and SEVIRI” (COCS) (Kox et al., 2011; Kox, 2012). It is very sensitive to optically thin cirrus clouds and was finally used to replace the original cirrus cloud mask algorithm.
COCS uses a neural network approach to convert the measurements in the infrared channels of SEVIRI into the two parameters ice optical thickness and cloud top pressure. The neural network was trained with a collocated dataset of SEVIRI observations and retrieval results for the “Cloud-Aerosol Lidar with Orthogonal Polarization” (CALIOP) aboard the polar orbiting satellite CALIPSO1. CALIOP is very sensitive to cirrus clouds and therefore the neural network delivers especially good results for thin clouds with τ < 1. However the effective radius, which has strong influence on the circumsolar radiation (see Sect. 3.6), is not retrieved. Furthermore, in an intercomparison of retrieved optical thickness between COCS and APICS, COCS showed a distinct bias to lower optical thickness for cirrus clouds for which APICS retrieved τ > 1. Because COCS is known to saturate at τ ≈ 2.5 (Kox, 2012), I assume that the optical thickness from APICS is more reliable for clouds with
τ > 1. For these reasons, the output from COCS was only used to generate a cloud mask for APICS.
Ostler (2011) and Bugliaro et al. (2012, 2013) computed the detection efficiency ηdet of MeCiDa utilizing airborne LIDAR observations. MeCiDa detects virtually all of the cirrus clouds withτ > 0.5 but only about 50% atτ ≈0.2. For most CST applications only clouds
32 3. Tools and Methods
with a slant path optical thicknessτs=τ /cos(θsun) smaller than 3.0 are relevant (withθsun being the sun zenith angle), otherwise too much light is extinguished in order to allow for energy production. Therefore, clouds withτ <0.5 account for a good share of the relevant clouds. The same studies show that COCS has advantages in detecting these clouds and reaches a detection efficiency of 80% even at τ = 0.15 while the detection efficiency of MeCiDa falls below 80% already betweenτ = 0.3 andτ < 0.4. It was therefore decided to derive a cloud mask from COCS results. All pixels that are assigned an optical thickness larger than 0.1 by COCS are assumed to be cloudy. This cut-off criterion of τ > 0.1 is necessary to keep the false alarm rate (FAR) at an acceptable level. Kox (2012) assessed the false alarm rate to be 5% for this cut-off using CALIOP observations as reference. The impact of the change from MeCiDa to COCS on derived circumsolar radiation is discussed in Sect. 5.2.1.
It should be kept in mind that there is no universally valid method to generate a cirrus cloud mask. Different methods differ in detection efficiency and false alarm rate. Often a higher detection efficiency goes along with increased false alarm rate. Furthermore these quality measures are subject to almost arbitrary definition: Consider only the treatment of satellite pixels that contain water and ice clouds at the same time. Should they be classified as cirrus cloud or not? This definition alone will strongly influence ηdet and the FAR of any algorithm. The approach in this study was to detect as many cirrus clouds as possible and at the same time to minimize the impact of false detections – which is possible only to a certain extent: The re-calibration of the SEVIRI measurements and the new albedo product help to minimize misinterpretation in clear sky conditions. However, they do not address the problems with water clouds (see also Sect. 4.2.2).
Because COCS uses SEVIRI’s infrared channels, it is independent of the sun’s position. The APICS cloud property retrieval however relies on reflected sun light, which causes deteriorations at sunrise and sunset as the plane parallel and the independent column approximation break down. Loeb et al. (1997) showed that for broken clouds with high vertical extent or for stratiform clouds with bumpy cloud tops the plane parallel approx- imation causes considerable errors even for sun zenith angles as small as 65◦. However, in a previous study Loeb and Davies (1997) concluded that the errors are in general less pronounced for optically thin clouds. In this study APICS results are used for sun zenith angles up to 80◦. This can be justified as the considered cirrus clouds are optically thin- ner and, as I assume, normally more homogeneous than the strato-cumulus like synthetic clouds investigated by Loeb et al. (1997).