Estimation method of spectral characteristic and area ratio of land cover based on probabilistic mixture model

Authors propose a parameter estimation method for the mixel problem in satellite images with coarse resolution. While conventional researches for the mixel problem have focused on only pure class probability distribution, Kitamoto and Takagi (1998) proposes a mixel decomposition method, which is based on an assumption that pure class has normal distribution, considers not only pure class but also mixel class probability distribution, and calculates mixel class distribution through convolution between pure class distributions. But the coarse resolution data from simulated data satisfying the assumption shows that the whole of pure class distribution does not obey one normal distribution. Therefore, in the first stage of our method, the original pure class distributions are restored based on the assumption that coarse pure class consists of two different normal distribution because some pure pixels are changed into mixels. In the second stage, the rest of pure class distributions and mixel class distribution are estimated. Both estimations in those two stages are conducted by EM algorithm. The result of the experiment using of simulation data show that our method can estimate each class distribution parameter more accurately than Kitamoto and Takagi´s method