Cloud detection with SVM technique

Following SPOT satellite family, PLEIADES-HR is a high resolution optical earth observation system developed by CNES and consisting in two satellites with a first launch by the end of 2009. Pleiades-HR will deliver both 70 cm Panchromatic images and 2.8 m multispectral (blue-green-red-near infrared) images with a 21km swath. As for SPOT images, Pleiades-HR ground segment includes a systematic and automatic cloud notation in order to feed the catalogue with a binary cloud mask and a confidence level notation used to monitor a manual re- notation in case of low confidence level. The cloud detection is operated from preview images generated through a fast decompression step of the low frequency wavelet subband : since Pleiades-HR compression is based upon a 3 levels wavelet transform, extracting the low frequency subband from the panchromatic and the multispectral images leads respectively to 5.6 m panchromatic and 22.4 multispectral images. In order to significantly improve the SPOT cloud notation process and get rid of frequent manual renotation, an innovative cloud detection technique has been studied and finally specified. It takes benefit from two kinds of information issued from the panchromatic and multispectral previews : multispectral reflectances and geometric parallaxes due to a panchromatic/multispectral stereoscopic angle that makes panchromatic/multispectral misregistration slightly sensitive to cloud altitudes and displacement velocity. These radiometric and geometric criteria are used as inputs of a Support Vector Machine (SVM) algorithm. SVM's are widely used for detection purposes and are based upon a training set of pixels with known criteria and classification (cloudy or cloud/free). Basically, the SVM technique aims to geometrically separate the training set represented in a Rn space, with n standing for the number of radiometric and geometric criteria taken into account for classification, using an hyperplane or some more complex surface if necessary. SVM training algorithm finds out the best frontier in order to maximize the margin, defined as a symmetric zone centered on the frontier with no training points included, and to minimize the number of wrong classification occurrences. In order to reach that goal, SVM training algorithm usually implements a Lagrangian minimization technique. Among SVM evens, one may quote the reduced complexity for the detection step : using an hyperplane to separate the two classes means that the detection process resumes to a scalar product in the Rn space for each pixel. Apart from the computation of radiometric and geometric criteria associated to each pixel to be classified, the complexity mainly lies in the training stage, executed only once for instance during the inflight commissioning period. Another advantage is the ability to generate a confidence mark for each pixel classification based upon the distance measured in the Rn space between the frontier and the point representative of the pixel to be classified : the general rule is that a large distance means a high confidence mark. The SVM technique has been implemented and assessed using a Quickbird Panchromatic/Multispectral data base, with various cloud coverages. Quickbird data are representative of future Pleiades images both in terms of resolution, spectral bands and even panchromatic/multispectral stereoscopic angle : it was thus possible to simulate quite representative panchromatic/multispectral quicklooks. The study also implied the generation of reference cloud masks that were produced through visual detection. A whole sensitivity analysis including the radiometric/geometric criteria taken into account and SVM parameters was conducted. Results were classically analyzed in terms of good detection and false alarm performances. The paper details the whole study of this SVM cloud detection technique within PLEIADES-HR framework.