Abstract: A bi-channel dynamic threshold algorithm is used to identify clouds on GMS-5 images. Threshold values of clouds and surface objects are gained in visible and infrared window channels by means of a statistic histogram analysis made of 32×32 pixel matrixes, followed by cloud recognition done from pixel to pixel each of the matrixes. The effect on cloud recognition of the chosen segment and bin size for dynamic smoothing in obtaining threshold values is concerned in terms of the histogram analysis. Results show that it is appropriate to choose 32×32 as segment size, thus ensuring a sufficiently large number of samples for each king in histograms clustering analysis; when the histogram is smoothed it is appropriate to choose the small bin at 1.2%(1.6K) for the visible (IR) channel so that a relatively short step length is applied in determining dynamic threshold value for required accuracy. Test of the results by visually examining images shows that with visible and IR data available the recognition precision from the algorithm is quite good at middle and low latitudes but the precision is not so good at higher latitudes where surface is colder, snow cover is large and the sun's elevated angle is low.