In order to realize cloud classification in all-time, all the spectral information of GMS-5 satellite imagery has been
exploited and made full use of in this paper. 2D~5D maximum likelihood algorithm was respectively used to
experimental research on cloud classification of multi-spectral GMS imagery. In contrast with 415 surface cloud
observation records in analysis region at 0800 local time, July 21,1998, if these cloud reports are strictly regarded as
true, the mean accuracy of 15 kinds of 2D~5D cloud classification results is 64.9%. After the similarities and
differences of satellite observation and surface cloud observation were surveyed, this paper points out that it is not
completely right and reasonable that the results of cloud classification are distinguished between right and wrong
absolutely according to surface cloud observation. Because the visual cloud observation from bottom to up on the
ground is inevitably unilateral, the two results of different observation is sometimes hard to compare directly,
contrasted the visual field observation from up to bottom of satellite. Therefore, the speciality of satellite observation
must be fully noted when cloud classification is achieved by using multi-spectral satellite imagery, so in this paper
the definition of distinguishing middle cloud and low cloud are put forward mainly according as brightness
temperature of cloud top and make full use of multispectral information to differ thin cirrus and thick cirrus from
low and middle cloud. To those samples classified as error by the criterion from surface cloud observation, it should
be reappraised based on the speciality of visual field observation from up to bottom and the actual situation of
satellite observation. The result reappraised to 35.1% of the "error" samples shows that 17.8% of those should be
thought reasonable. The mean accuracy of 15 kinds of 2D~5D cloud classification results has been to 82.7% and the
maximum accuracy is up to 87.0%, which is obtained from the 4-D maximum likelihood dynamic clustering of four
wave band (IR, VIS, WV and TIR2-IR1 ) GMS imagery data. The accuracy of cloud classification also reaches to
81.4% using the other four band (IR, WV , T WV -IR1 and TIR2-IR1) imagery , especially when there is no VIS imagery at
night.
The final example shows on condition that multispectral information has been fully used, different spectral bands
combination are utilized reasonably day and night respectively, the reasonable cloud classification will be well
realized in all-time by using maximum likelihood algorithm.
|