Research On Object - Oriented Classification Of High - Resolution Image Based On Texture Analysis

With the rapid development of spatial resolution in satellite remotely sensed data, thetraditional pixel-based image processing methods encounter bottleneck for the inability to makefull use of detail spatial information in high spatial resolution images. In addition, the classificationresults derived from the pixel-based analysis always have many misclassifications and errorclassifications, due to the matter of ’same material with different spectral’ and ’same spectral fromdifferent materials’. Moreover, there would be serious “salt-and-pepper noise” in the pixel-basedclassification results, which could drastically affect the classification accuracy. The object-orientedclassification method,however, can effectively avoid the abovementioned problems. This methodhas received extensive attention and been applied in many studies.Texture featureshows important spatial feature information of images, it has been widely usedin remotely sensed image classification. Many previous studiesapplied texture feature assupplemental information and got good classification results.But most of the previous studies havefocused on the traditional pixel-based classificationmethod, and the texture studies by object-oriented classification with single scale. The influence of texture feature on classification accuracyderived from object-oriented classification with multi scales has not been deeply studied. Basedon the previous studies, we chose the Fengnan district in Tangshan city as our study area andgenerated the multi-scale object-oriented classification system.Through the ESP(Estimation of Scale Parameters) tool and many experiments, we decided theoptimal segmentation scale and parameter for the dominant land cover types in the study area. Themulti-scale segmentation hierarchy (81,47,16) with three levels was built, and the multi-levelclassification system was established. It showed the advantage of multi-scale segmentation inobject-oriented classification.The study extracted8gray-level co-occurrence matrix (GLCM) texture features and3localspatial statistic (LSS) texture features. The classifiers of SVM(support vector machine) andNN(nearest neighbor) were used to study the influence of different texture features on overallaccuracy (OA), producer’s accuracy (PA) and user’s accuracy (UA). The experiment resultsshowed that the OA and PA and UA of most classified classes could be significantly improved byadding even one texture feature. The added texture information had greater effect on OA of NNclassification results, while relatively smaller effect with around1%improvement of SVM’s OA.Classification results with the Geary’s C texture feature had the best OA (79.06%). Compared withclassification results generated from multiple spectral bands without texture feature, the adding ofGeary’s C texture feature improved4.6%of OA. We chose some texture features to study the influence of texture scale parameter on classification accuracy. The results showed that the changeof scale parameter affected classification results, but the influence could be reduced by the self-mechanism of object-oriented classification.This study proposed a method to select the best feature combinations, based on ant colonyalgorithm. It ensures the great reduce of the feature dimensions, under the condition of highclassification accuracy. In addition, the method can obtain the best feature combination fromsamples without classifying. In this study, we obtained the best texture feature combinations forSVM and NN classifiers in the study area and validated the results.