Articles | Volume 10, issue 11
Research article
14 Nov 2017
Research article |  | 14 Nov 2017

Characterisation of the artificial neural network CiPS for cirrus cloud remote sensing with MSG/SEVIRI

Johan Strandgren, Jennifer Fricker, and Luca Bugliaro

Abstract. Cirrus clouds remain one of the key uncertainties in atmospheric research. To better understand the properties and physical processes of cirrus clouds, accurate large-scale observations from satellites are required. Artificial neural networks (ANNs) have proved to be a useful tool for cirrus cloud remote sensing. Since physics is not modelled explicitly in ANNs, a thorough characterisation of the networks is necessary.

In this paper the CiPS (Cirrus Properties from SEVIRI) algorithm is characterised using the space-borne lidar CALIOP. CiPS is composed of a set of ANNs for the cirrus cloud detection, opacity identification and the corresponding cloud top height, ice optical thickness and ice water path retrieval from the imager SEVIRI aboard the geostationary Meteosat Second Generation satellites. First, the retrieval accuracy is characterised with respect to different land surface types. The retrieval works best over water and vegetated surfaces, whereas a surface covered by permanent snow and ice or barren reduces the cirrus detection ability and increases the retrieval errors for the ice optical thickness and ice water path if the cirrus cloud is thin (optical thickness less than approx. 0.3). Second, the retrieval accuracy is characterised with respect to the vertical arrangement of liquid, ice clouds and aerosol layers as derived from CALIOP lidar data. The CiPS retrievals show little interference from liquid water clouds and aerosol layers below an observed cirrus cloud. A liquid water cloud vertically close or adjacent to the cirrus clearly increases the average retrieval errors for the optical thickness and ice water path, respectively, only for thin cirrus clouds with an optical thickness below 0.3 or ice water path below 5.0 g m−2. For the cloud top height retrieval, only aerosol layers affect the retrieval error, with an increased positive bias when the cirrus is at low altitudes. Third, the CiPS retrieval error is characterised with respect to the properties of the investigated cirrus cloud (ice optical thickness and cloud top height). On average CiPS can retrieve the cirrus cloud top height with a relative error around 8 % and no bias and the ice optical thickness with a relative error around 50 % and bias around ±10 % for the most common combinations of cloud top height and ice optical thickness. Similarities with physically based retrieval methods are evident, which implies that even though the retrieval methods differ in the implementation of physics in the model, the retrievals behave similarly due to physical constraints. Finally, we also show that the ANN retrievals have a low sensitivity to radiometric noise in the SEVIRI observations. For optical thickness and ice water path the relative uncertainty due to noise is less than 10 % down to sub-visual cirrus. For the cloud top height retrieval the uncertainty due to noise is around 100 m for all cloud top heights.

Short summary
We characterise the the performance of a set of artificial neural networks used for the remote sensing of cirrus clouds from the geostationary Meteosat Second Generation satellites. The retrievals show little interference with the underlying land surface type as well as with possible liquid water clouds or aerosol layers below the cirrus cloud. We also characterise the retrievals as a funtion of optical thickness and top height and gain better understanding of the retrival uncertainties of CiPS