Multi-Sensor Cloud Screening and Validation: IdePix and PixBox
Brockmann, Carsten; Paperin, Michael; Danne, Olaf; Ruescas, Ana
Brockmann Consult GmbH, GERMANY

Processing of Level 1 data to Level 2 data requires the characterisation of each individual pixel into either land, water or totally cloudy. Further characterisation information is required in order to apply specific correction algorithms, or to configure the higher level processing. Such characterisation includes: presence of cirrus clouds, cloud shadow, spatially mixed cloud-land or water-land, coastline, submerged vegetation in water pixels, just to name some important cases.

Started in the framework of the ESA DUE CoastColour and Globalbedo projects, and further developed in the Ocean Colour and Land Cover CCI projects, an approach has been developed to apply an extensive pixel characterisation at Level 1 and to amend the product with additional pixel characterisation information. This includes the final information as a binary flag, but also floating number, e.g. a cloudiness index which gives the user the possibility to apply his/her own atmospheric correction depending on this index.

The underlying pixel characterisation methodology is a sensor independent feature based approach (Identification of Pixels, IdePix). Feature values, such as brightness, whiteness, temperature, etc. are defined and are calculated according to the specific sensor IdePix is applied to. Feature values are normalised to 1, and arithmetic is applied to estimate the final cloudiness of a pixel depending on the available individual feature values. The main purpose of using IdePix is cloud screening; however, other characteristics such as cloud shadow, coastline or mixed pixel type can also be derived. The IdePix methodology has been applied to MERIS, AATSR and SPOT-VEGETATION, and is currently extended to MODIS, VIIRS and OCM. Extension to new sensors are under discussion.

Validation of pixel characterisation is a critical but very difficult task. We have developed a methodology to compare the IdePix output to a large database of manually classified image pixels. The classes include clear land, ocean and cloud, but also ambiguous cases such as for example partly cloudy or further characterised pixels (e.g. inland water) are also included. This database is called PixBox, and was developed for MERIS Reduced Resolution data (contains 110 000 manually classified pixels), MERIS Full Resolution data (20 000 pixels), has been extended to SPOT-VGT (20 000 pixels) and will be completed for AATSR during 2012. We applied PixBox to evaluate the performance of the cloud screening of MERIS 3rd reprocessing, and of LandCover and Ocean Colour CCI. PixBox will also be collected for Sentinel 2 and 3 and the method has been proposed to the Sentinel 3 Validation Team call.