Several years of studies to learn spectral signatures of clouds and cloud-free surface types at SMHI (a supervised training phase) resulted in the development of a prototype for an operational automated cloud classification model (Karlsson, 1989). The final classification model, named SCANDIA (the SMHI Cloud ANalysis model using DIgital AVHRR data) was then implemented in 1988. After a few updates, it has been kept unchanged since 1991, although some modifications have been introduced in parallel (see below).
In the following, only a brief description of the main characteristics of the SCANDIA model is given. A more thourough description can be found elsewhere (Karlsson and Liljas, 1990; Karlsson, 1996).
SCANDIA is a supervised thresholding model, where typical class domains are defined by hyperboxes in a seven-dimensional feature space. To allow for the natural elongated and skewed pixel distributions in feature space, each cloud or surface class is often described by several boxes.
The classification model makes use of calibrated and geometrically transformed (using polar stereographic projection) imagery from all five AVHRR/2 channels at maximum horizontal resolution (at nadir 1.1 km).
In the first version of SCANDIA which went operational in 1988 AVHRR scenes are classified by using seven image features for two different areas covering the entire Sweden and adjacent areas. One feature is a land/sea mask based on a geographical map, whereas the other six are the following pure spectral features, based on all five AVHRR channels:
- CH1 Bi-directional visible reflectance (not corrected for sun elevation)
- CH1-CH2 VIS-NIR reflectance difference
- CH3-CH4 Brightness temperature difference
- CH4 Brightness temperature
- CH5-CH4 Brightness temperature difference
- TEX4 Local brightness temperature variance (based on highpass filtering with a 5x5 pixel window) in AVHRR channel 4.
Variations in spectral features due to a varying sun elevation is compensated for by using thresholds depending on the sun elevation according to 12 pre-defined categories (including night time conditions).
The pixels in each AVHRR scene are separated into a maximum of 57 classes (cloud types and surfaces) depending on the sun elevation. However, the final image product as presented to the forecaster never comprises more than 23 classes.
Examples of resulting cloud classifications are shown below for areas covering northern (right image) and southern (left image) Sweden. The images are produced fof a NOAA 14 pass at 12:31 UTC August 4, 1998. Class colurs are explained further down.
The feature CH3-CH4 is central for the classifier. The thresholds in this feature for the two categories water clouds and ice clouds differ at night and at very low sun elevations, whereas during daytime the same threshold is used. However, a special treatment of sunglints is necessary to avoid confusion with low-level cloud types. A combination of features CH1 and CH3-CH4 is then used based on differences in reflectance for clouds and sunglints. For the final determination of cloud types, several threshold tests follow based on all features (see Karlsson, 1996).
A slightly modified version of the model was implemented in 1994, executed in parallel with the original SCANDIA algorithm. It is operated on a larger northern European area but with reduced horizontal resolution (4 km). A more consistent treatment of sun elevation dependence has been included and, in addition, temperature information from short-range NWP model forecast (HIRLAM 9- or 12-hour forecasts) is utilised. Especially, the use of forecasted surface temperature is implemented to improve cloud classifications at dawn/dusk and in problematic situations at night (i.e., situations with thin Cirrus clouds superposed over low clouds).
The small scale texture feature used in the earlier version is omitted in this modified version of SCANDIA since it is not applicable in coarse resolution images.
An example of the result of the 4 km classification is shown here. A colour legend describing the cloud and surface types is displayed to the right. The data are from the NOAA 14 satellite pass at 12:31 UTC August 4, 1998.
Specific user-oriented features have also been introduced to the classification result to facilitate the use in weather forecasting. Problematic classification conditions are identified (e.g., situations with strong low-level temperature inversions or with shadows at very low sun elevations etc.) and the forecaster is alerted by use of unclassified classes in classification images. This product is described by Karlsson (1996).
Adaptation to variable conditions
A unique set of thresholds is defined for each of several pre-defined categories related to existing illumination conditions, weather types and satellites. The season (one of the set "Summer, Spring, Winter, Autumn"), the current sun elevation (determined by one of 12 defined sun elevation intervals) and the satellite (here NOAA-11, NOAA-12, NOAA-14 and recently also NOAA-15) specify which category and thus which set of thresholds that is used to classify the scene in each of two defined geographical areas. A dependence on the present weather type is introduced by use of NWP model analyses (temperatures in pressure levels 700 hPa and 500 hPa).
Treatment of sub-pixel cloudiness
All thresholding models are based on the assumption that clouds are found when radiances are significantly deviating from expected radiances from cloud-free land or ocean surfaces. Typically, cloudy pixels appear brighter than a threshold in visible channels and colder than a threshold in infrared channels. This means that threshold radiances cannot be perfectly the same as cloud-free radiances and thus, some cloudiness will always be missed or underestimated by thresholding schemes provided that clear radiances are correctly estimated.
SCANDIA includes treatment of a special class, called "sub-pixel clouds", that contains pixels being judged as cloud contaminated and thus not totally cloud-filled. Radiances are here close to cloud-free radiances but these pixels have to be identified and discarded in applications aimed at determining surface properties (e.g. sea surface temperature retrievals). However, no estimation of the fractional cloud cover within individual pixels is made by SCANDIA. The reason is the difficulty to separate radiances emerging from a pixel filled with an optically very thin cloud from radiances emerging from a pixel marginally filled with an optically very thick cloud element.
Karlsson, K.-G., 1989: Development of an operational cloud classification model. Int. J. Remote. Sensing, 10, 687-693.
Karlsson, K.-G. and Liljas, E., 1990: The SMHI model for cloud and precipitation analysis from multispectral AVHRR data. PROMIS Report 10, SMHI, 74 pp.
Karlsson, K.-G., 1994: Satellite-estimated cloudiness from NOAA AVHRR data in the Nordic area during 1993. Reports Meteorology and Climatology, SMHI, 51 pp.
Karlsson, K.-G., 1995: Estimation of cloudiness at high latitudes from multispectral satellite measurements. AMBIO, 24, 33-40.
Karlsson, K.-G., 1996a: Cloud classifications with the SCANDIA model, Reports Meteorology and Climatology, SMHI, 36 pp.
Karlsson, K.-G., 1996b: Validation of modelled cloudiness using satellite-estimated cloud climatologies, Tellus, 48A, No. 5, containing papers from the First Study Conference on BALTEX, Visby, Sweden, 28 Aug.-1 Sep. 1995, 767-785.
Karlsson, K.-G., 1997: Cloud climate investigations in the Nordic region using NOAA AVHRR data, Theor. Appl. Climatology, , 57, 181-195.
Karlsson, K.-G., 1999: Satellite Sensing techniques and applications for the purposes of BALTEX, Contr. Atm. Phys., in press.