MERIS-AATSR Synergy for Cloud Detection
Lydwine Gross-Colzy(1), Mickael Ferrand(1) and Patrice Henry(2)
(1) Capgemini Sud, 15 avenue du docteur Grynfogel, 31036 Toulouse Cedex 1, France
(2) Centre National d'Etudes Spatiales, 18 avenue Edouard Belin, 31400 Toulouse, France
In this paper, we demonstrate the feasibility of the MERIS-AATSR quasi instantaneous merging for the particular problem of cloud detection. Cloud detection is an important step of earth observation data processing, which goal is to retrieve properly atmospheric and surface parameters. Indeed, measurements contaminated by clouds are source of significant errors in atmospheric, oceanic and land cover parameters retrieval, and thus may lead to erroneous forecasts.
Up to now, most cloud detection methods are based on an analysis of their radiative properties: the clouds are cold, dense and bright while the surfaces of earth and ocean are warmer and darker respectively. Spectrally dependant contrast differences generally provide satisfactory results for most observational geophysic conditions, and thus technical thresholds applied to measurements in the visible, near-infrared and thermal infrared are largely used for detection of clouds. But the drawback of using thresholds cascades is that the separation between partly-cloudly scenes and clear scenes is not linear in the measurement space, which results in a non-negligible amount of false classifications. Moreover, for sensors like MERIS which measure in the visible and near-infrared part of the spectrum but which do not have information in the short-wavelength, mid-wavelength or in the thermal infrared regions, some undeterminations remain which complicate the separation of clouds from bright deserts, or clouds from snow.
We illustrate in this paper, using an appropriate technique for merging MERIS and AATSR L1B radiometric measurements, the results of a non linear and fast method developed to detect and classify clouds. The methodology is generic, and may be applied from one to several sensors from diverse nature. We first show the results obtained on MERIS measurements alone, and compare the results to the BEAM cloud probability processor on difficult targets like bright deserts in the saharian region and and snow at high latitude regions. Secondly, we show the results obtained on MERIS and AATSR merged measurements from the same day, and compare them to previous experiments results and the AATSR cloud flag. Our classifier is applied either over land or oceanic surfaces and for this study we chose to classify clouds within three labels: cloudy, partly cloudy over ocean, partly cloudy over land.
The classifier we use is a labelled topological neural network (or Kohonen map) containing a reasonable amount of referent vectors. The referent vectors are first calibrated using the non-supervized Kohonen algorithm on a large data set containing spectral information (and more generally merged spectral information) from various types of observations (clouds, different land cover and ocean surfaces). The referent vectors are then labelled with the help of an expert, and eventually referent vectors which have similar radiometric properties, and hence which are neighbours on the topological map, are identified by the same label. Once calibrated and labelled, the classifier may be applied to series of images containing various landscapes, cloud types etc. Because we label the referent vectors and not the data themselves, this method do not require a large amount of labelled data. Moreover, the frontiers we build on the topological map between cloudy scenes and non cloudy scenes may be non-linear.
Thirteen out to fifteen MERIS spectral bands are used (the absorption bands located in the oxygen and water vapour absorption bands are evicted) and we use either visible to short-wavelength AATSR reflectance or AATSR brightness temperatures. In order to enhance discrimination of clouds from the rest and to allow quasi instantaneous merging, we encoded the spectral signature of each sensor on several different mathematical space. We use PCA and CIE Lab encoding to extract discriminative spectral features from the mesurements, letting aside intensity information in order to avoid directionality and seasonal issues.
Keywords: ESA European
Space Agency - Agence spatiale europeenne,
observation de la terre, earth observation,
satellite remote sensing,
teledetection, geophysique, altimetrie, radar,
chimique atmospherique, geophysics, altimetry, radar,