Recognition Research On Texture Feature In Remote Sensing Image

With the development of space remote sensing technology and widely used high-resolution remote sensing images, the image depending only on the characteristics of spectral information as a feature has great limitations. Texture as a reflection of spatial information having become important features of marked ground objects, texture feature extraction methods and how to use the features of object recognition has been the focus of remote sensing image processing is of great significance.Most previous studies for the recognition of spectral information is not very good at binding non-spectral information, this paper will be in accordance with the spectrum of information in Jinan ETM image is divided into two types of Feature. Ground objects by vegetation, mountain forest tree, shrub having similar spectrum can be used as a class, residential area without vegetation cover and bare soil, building a class, follow-up study only around the each class of spectra.Gray Level Co-occurrence Matrix is an effective way for texture analysis, texture feature vector composed by contrast, correlation, energy, entropy, inverse difference moment. In this paper, texture features extracted from TM1, TM2, TM3 band of ETM by GLCM , parts of them are used to construct gray co-occurrence vector for each type of spectral classification. Land, mountain forest tree, shrub land use {ASMB3, ASMB2, CORB1, ASMB1, ENTB1, IDMB1}.Residential construction and the bare soil use {CONB3, ASMB3, IDMB3, CONB2, ASMB2, ENTB2, IDMB2, CONB1, CORB1}.GLCM describes texture features from pixel correlation of gray and the fractal dimension reflects the structure of self-similar. In this paper, using differential box counting method to calculate the dimension (DIM) of Ground objects from TM1, TM2, TM3 band ETM image and the dimension of the texture features as a feature and then with the combination of Co-occurrence of the vector form the gray co - dimension feature vector. Land, mountain forest tree, shrub land use {ASMB3, ASMB2, CORB1, ASMB1, ENTB1, IDMB1,DIMB3, DIMB2, DIMB1}.Residential construction and the bare soil use {CONB3, ASMB3, IDMB3, CONB2, ASMB2, ENTB2, IDMB2, CONB1, CORB1, DIMB3, DIMB2, DIMB1}.BP neural network and Bayesian network are mature technology used in pattern recognition. In this article gray co-occurrence vector and co-dimension feature vector extracted from ETM images of Jinan are used in recognition experiment with BP neural network and Bayesian network. The experiment demonstrates feasibility of the method of using co-occurrence texture feature vector, Discusses pros and cons of co-occurrence texture feature vector and gray co-dimension feature vector, Compares BP Neural Network with Bayesian Network.Experimental results show that the proposed construction of gray co-occurrence vector is feasible. Ground objects can be effectively recognized by gray co-occurrence vector and gray co - dimension feature vector with BP neural network and Bayesian network, recognition rate of 70%. Application of gray co-dimension feature vector is better than the gray co-occurrence vector. BP neural network's recognition accuracy and time consumption all more than Bayesian network, the accuracy 2% -5% higher. The stability of object types is one bottleneck in BP, leading to its robustness less than naive Bayesian network.