Towards the operational cloud classification in Finland Otto Hyvärinen COST 722 Expert Meeting, Helsinki, 9 June 2004
Contents Cloud classification work with NOAA/AVHRR –in FMI –in Nowcasting SAF (PPS) Problems Future plans
FMI approach to cloud classification with NOAA/AVHRR The statistical pattern recognition approach –Collect a lot of training data –Approximate posterior probabilities with neural networks –Use a principled way of handling the uncertainty No physics, only data! –Different models for day, night and twilight Unfortunately, the development has stalled...
20 September 2001, 5:54 UTC
Cloud Classification
Goodness of Cloud/Surface Separation
Nowcasting SAF approach with NOAA/AVHRR Developed mostly in SMHI Traditional thresholding –Thresholds computed with the help of the radiative transfer models (RTTOV) Uncertainty handled more ad hoc
An example from 8 June 2004, 06:20 UTC Cloud MaskCloud Classification
How do they compare? Comparison against SYNOP observations of total cloudiness All stationsStations from "the training area"
Problems? Common problems –Twilight! –AVHRR FMI problems – no physics should explore RTMs Nowcasting SAF problems – bad decision making methods should explore pattern recognition methods
Towards future Is Nowcasting SAF method for AVHRR "good enough" for us? From AVHRR to MODIS, VIIRS, etc –how to make this as least painful as possible? –SEVIRI, how useful is it in Finland, really? From the satellite cloud classification to the cloud analysis –how to combine remote sensing and in-situ observations with the model data to real 3D analysis of clouds, fog, and visibility?