Object-oriented Classification Techniques For Remote Sensing Image Based On Genetic Algorithms

Remote sensing image, especially high-resolution remote sensing image, has rich spatial information, such as geometry, structure, texture etc. Traditional pixel-based classification method, which utilizes only spectral information, not only leads to the wastage of spatial information, but also affects the effectiveness of classification seriously, because of the phenomenon of same objects having different spectral character and different objects having same spectral character and the existing of noises. Object-oriented classification techniques of remote sensing image, integrating information of spectrum, shape and texture of image, utilizes spatial information to a great extent so that to meet the drawback of the traditional method while dealing with classification of remote sensing image.Object-oriented classification needs to use a variety of high-dimensional features with significant differences in nature. Therefore, a single classifier can not achieve a stably ideal result while dealing with those features. Classifier fusion techniques of multiple classifier system have been a hot topic in recent years. In response to the characteristic mentioned above, classifier fusion based on feature selection techniques should be a possible solution.Based on the above characteristic, firstly genetic algorithms, which have been widely used in optimization problems, was applied to feature selection techniques, and then a multiple classifier system model was designed as extension of genetic-algorithm-based feature selection techniques. As there are correlations among features, a new multiple classifier system model was constructed with disjoint feature subspace constraint. These two models were trained by supervised methods and oriented by improving the classification accuracy. The construction of multiple classifier system is completed by genetic algorithms, which determine what optimized feature subspace each sub-classifier should uses.Experiments show that two models, especially the latter, could effectively improve classification accuracy of object-oriented remote sensing image with high-dimensional features.