The Study Of RS Image Classification Methods Based On Feature And Its Applications

This dissertation focuses on the image classification utilizing the texture features of SPOT5 panchromatic image and spectrum features of SPOT5 multi-spectral image, and realizes the dynamitic monitoring in land use based on these two kinds of features.The main study contents of this dissertation are as following:⑴Image classification based on texture feature: Establishes the texture feature database with gray-level histogram, gray co-occurrence matrix and gray-gradient co-occurrence matrix. And 22 texture features are used in feature extraction and selection experiments. Then does the experiments separately by transformation based on classifiability criterion, discrete K-L transformation, Fisher transformation, direct choosing methods on feature extraction and selection. And compares and analyzes the precision of these different classification methods under confusion matrix.⑵Image classification based on spectrum feature: Proposes the Spectral Curve Feature Model and Tree-judgment Model. Then compares the results of these two classification methods with Maximum Likelihood Classification, K-L transformation, Spectral Angle Mapper, spectrum and texture feature, and Erdas software, and proves the precisions and feasibilities of these new methods.⑶Introduces the Maximum Likelihood Classification in dynamitic monitoring in land use based on spectrum feature, proposes the Class Statistic and Judge Function, and improves the precision of monitoring. Then applies the feature extraction and selection in dynamitic monitoring in land use based on texture feature. And makes the compare, analysis and precision estimation of the above two monitoring methods.