Satellite Image Scene Classification Using Spatial Information

In the study of high resolution satellite image classification, the images show twocharacteristics: First, with the improvement of spatial resolution, the objects showmore details in a satellite scene, such as structure, texture and color. Second, objectsmay appear at different orientations and scales in the same category scene of HRsatellite images and different scenes may contain the same object. Meanwhile, thebrightness of the same scene is influenced by lighting under different weatherconditions.The above two characteristics put big obstacles to represent images for sceneclassification. Image representation plays a key role in scene classification. Unlikelow-resolution satellite images, which can be effectively described through textureand intensity features, these simple features cannot fully express the objectsinformation for each HR satellite image owing to the characteristics. In order toenhance the local feature representation capacity and improve the classificationperformance of high resolution satellite images, we should employ some properties orrelations for describing every image scene.This research includes next three main tasks. The first is the feature selection forhigh resolution satellite images, which plays a key role in scene classification. Asuitable feature could be good at describing some prominent regions, such as sky,road and river. Scale-invariant feature transform (SIFT) features have some invariantproperties, such as affine, rotation, brightness and scale. Considering the abovementioned characteristics of high resolution satellite images, therefore, SIFT isnecessary in our application.The second is constructing visual dictionary after feature selection. There are twoschemes that can be implemented to extract features for constructing visual dictionary.This research adopts one of scheme for better generalization ability. This scheme isnot only suitable for the randomly selected samples for constructing the visualdictionary in original sample library, but also suitable for constructing visualdictionary in the new sample library. It makes the experimental result more convincedand reliable.The third is employing spatial information of features to enhance the ability of representing images. The spatial pyramid matching model (SPMM）is introduced forspatial coding and matching for HR satellite images. To improve the classificationperformance, through spatial pyramid modeling, the spatial information of localfeatures is used for image classification.The experiment is based on MATLAB platform. The LIBSVM toolbox is used totrain an SVM classifier on the training set. The average classification accuracy is82.6%. The greater the distinctiveness of different scene is, the higher theclassification accuracy is. In the comparison of standard deviations for several groupsexperimental results, it is demonstrated that the results of each experiment arerelatively stable and are less affected by choosing training set and testing set randomly,thus with a robustness. We conclude that, by utilizing the spatial information of localfeatures through SPMM encoding, our approach can enhance the ability of describingHR satellite images and improve the classification performance.