Gaussian Mixture Model of Texture for Extracting Residential Area from High-resolution Remotely Sensed Imagery

Using high-resolution remotely sensed imagery to timely detect distribution and expansion of residential area is one of most important jobs of national 1:5 spatial database updating. In view of complicated spatial characters of residential area and working disable of current automatic interpretation methods based on spectral features on high-resolution remotely sensed imagery, a classifier based on Gaussian Mixture Model (GMM) of texture is proposed. The combination of co-occurrence texture features (contrast, entropy, mean, standard deviation and correlation included) and edge density are used to substitute for spectral features as classification features. A mixture density function is employed to represent classes’ distribution in texture spaces. And residential areas are extracted through the classification based on GMMs obtained through estimating using Expectation Maximum (EM) algorithm. The proposed method is examined by an IKONOS panchromatic imagery.