``Mixed pixel'' or ``mixel'' denotes a pixel in which multiple classification classes are contained; and the area of each class within a mixel is called ``area proportion.'' This paper proposes a method for estimating the expected area proportions of mixels based on the probability density function (pdf) of each class, which is also estimated from the image histogram by mixture density estimation. Here the most significant feature of the model used in this paper is the introduction of a new type of stochastic model called ``mixel density.'' This new model, with flat shape, works effective in estimating the parameters of pdfs from long-tail image histogram caused by the presence of many mixels on an image. Experiments of pixel-based classification applied to satellite imagery demonstrates that our mixture density model shows better fitness than a model without mixel densities, which results are then quantitatively evaluated using Akaike information criterion.

      @Article{ kt:ieicetrans98-a,
      author = {Asanobu KITAMOTO and Mikio TAKAGI},
      title = {Estimating the Area Proportions of Mixels Using Mixture Density Estimation with Mixel Densities},
      journal = {Transactions of Institute of Electronics, Information, and Communication Engineers (IEICE)},
      volume = {J81-D-II},
      number = {6},
      pages = {1160-1172},
      year = 1998,
      month = 6,
      doi = {},
      note = { (in Japanese)},
      }