Research on remote sensing image classification based on Neuro-Fuzzy and texture analysis

Classification accuracy is one of major factors influencing the application of classified image. This paper introduces the Neuro-Fuzzy system as a classifier and investigates the performance of this method in image classification. Using Landsat ETM+ satellite image as source data, gray Level Co-Occurrence Matrix (GLCM) is calculated to characterize the texture characteristics of the surface objects. Five texture features including Entropy, Contrast, Correlation, Inverse Difference Moment, and Angular Second Moment are derived from GLCM and used as inputs of Neuro-fuzzy classifier to extract the target surface objects from the Landsat ETM+ image. The results indicate that all target objects including water, mountain, gobi, vegetation, desert and resident area could be well separated from each other based on texture characteristics, and the Neuro-fuzzy based classification method can get better classification results with overall accuracy of 78.3% compared to commonly used method such as maximum likelihood classification.