A new combined nonparametric algorithm based on a four-step procedure is developed for segmentation of multi-spectral space-made photographs of the Earth's underlying surface and cloudiness. At the first step, fragment-by-fragment local clustering of video data is performed by use of the Bhattacharyya distance or Kullback divergence. At the second step, adjacent obtained classes are unified by use of the empirical risk functional. At the third step, enlarged classes serve as a learning material for a nonparametric algorithm of pattern recognition. Finally, at the fourth step, the pattern recognition algorithm performs segmentation of the whole image. This approach makes it possible to solve the problem of compromise between awkwardness of the initial data and necessity to use adequate models of images under recognition based on nonparametric estimates of unknown conditional probability distributions. Besides, the problem of studying complexes of features for information content is being solved in the sense of the minimum of the empirical risk criterion.