Feature Selection For Object-oriented Classification Of High Resolution Remote Sensing Images

Object-oriented image classification approaches, mainly including the nearestneighbor(NN) algorithm and membership function based on fuzzy, is a suitable for highspatial resolution image information extraction technology. It is chiefly characterized by smallsample and high dimensionality of features. And the membership function of fuzzyclassification is more suitable for the larger image area. However, in the actual classificationof membership function, according to the features of remote sensing images and experienceknowledge, a small amount of features by analysis and try were used to classification. Finally,the classification result was good or not to determine whether the features would be selected.Obviously, it would cause subjectivity and blindness. And it isn’t scientific and practical.Therefore, selected features could not ensure the classification speed and accuracy. Thisresearch was founded by―Research on the land use type quickly extraction with highresolution data"(No: E0202/1112/0104) in the―12thFive Year Plan‖. For the feature selectionof object-oriented classification using membership function, this research mainly develops thefollowing aspect to study:1. It was proposed that the feature selection algorithm to remove Irrelevant features.Based on small sample size, the Relief algorithms which is considered one of the best way inthe feature selection area, were improved from aspects of drawing sample randomly, KNNdenoising, iteration times in this paper. And25dimensions feature were selected from initial63dimensions feature based on the improved Relief algorithm during the experiment in thepaper. The results showed that the improved Relief algorithm is more useful for featureselection in the object-oriented image classification,―dimension disaster‖is avoidedefficiently, and based on the algorithm provides high quality Relevant Features for thefollow-up of J-M distance. Therefore, the improved algorithm has two virtues, strong abilityof removing irrelevant features and without restriction of data type;2. A compositional feature selection algorithm for Object-oriented Classification was proposed. It is based on combination of improved Relief and J-M distance. The principle ofalgorithm is that using J-M distance calculated the separation between different classes basedon relevant features to select the features of classification for each land use type. The resultsshowed that it is easier selected feature for each type of terrain based on this algorithm. Andthe efficiency of classification based on membership function was increased. Whether theview of the classification precision or efficiency, the validity of the feature selection algorithmwas proved in the paper. The proposed feature selection algorithm can choose featureseffectively, narrow range of features and shorten the classification of time. So it has some ofthe research significance and application value;3. The design and implementation of feature selection module for object-orientedclassification. For verifying the validity and correctness of the method, the feature selectionalgorithm was designed and implemented based on Microsoft.NET Framework2.0/COMtechnology.