We propose a new approach for detecting edges based on Contourlet probability distribution. First we analyze the mixture of Gaussian distribution traits of the Contourlet coefficients. Then we establish the probability model regarding the Contourlet coefficients which can be described by the big state and the small state. At the same time

we separate the linear singular signals on the model of the image. Afterwards

we improve the maximal between-class variance by a threshold selection method based on the variance between different classes and the in-class variance. This can ensure the maximum distance between classes

while simultaneously increasing the degree of polymerization in a class. Furthermore

we use the threshold for the binarization of the separated singular signals and extract the edge message. Compared to traditional methods

our method detects smooth edges in remote sensing images effectively. Meanwhile

it can shield the unimportant singular information

making it is useful for practical applications.