The Multiscale Classification Method Of High Resolution Remote Sensing Images

In recent years, high resolution remote sensing images are widely used in variousfields, including not only traditional departments like territorial resources, geologicalsurvey and mapping, but also the fields of urban planning, transportation, tourism, andecological environment. However, how to accurately extract and effectively utilize theinformation from a huge number of high resolution remote sensing images remains tobe one of major challenges. Due to the complexity and multi-scale property of highresolution remote sensing images, the traditional pixel-oriented method has manyshortcomes. The object-oriented image classification method can overcome theshortcomes and it is gradually becoming a main stream in this field.In object-oriented classification method, images are firstly segmented into severalmeaningful objects, feature extraction and classification are conducted in the next step.This paper paid attention to the key points of object-oriented image classificationmethod. The main contents are presented as follows:(1) Make an initial segmentation on high resolution remote sensing images withthe developed multi-scale segmentation algorithm based on watershed transformation.First, an anisotropic diffusion filter was used for pre-processing. Second, the multi-scalemorphological gradient feature was extracted. Then the marker-based watershedalgorithm was carried out. Finally multi-scale region merging based on theobject-oriented spectrum and heterogeneity shape indicators would be employed.Having considered the complexity of multi-spectral and multi-scale of high resolutionremote sensing images, the presented method, to some extent, improved imagessegmentation results.(2) Introducing the Bag-of-Words model from text analysis to express the featureof remote sensing images. It crossed the "semantic gap" by means of changing low-levelvisual features into high-level semantic ones. The paper took the scale characteristics ofremote sensing image into fully consideration and built multi-scale visual words basedon a Gaussian scale pyramid. Experiments showed that the new method was better thanclassical classification methods. (3) Finally, on the basis of Bag-of-Words model, probabilistic latent semanticanalysis method was adopted to analyze the image objects because of its strong abilityof identifying synonyms and polysemy. This method is non-supervised and does notneed training samples. In order to verify the effectiveness of the method in this paper, ahigh resolution aerial remote sensing image, a Quickbird image and an IKONOS imagewere selected for classification experiments. Experiments showed that the proposedmethod was superior to other non-supervised classification methods, and was slightlybetter than the supervised object-based classification method (taking airborne remotesensing images for example, the overall classification accuracy of the proposed methodwas90.57%, the pixel-oriented ISODOTA clustering method was42.45%, thepixel-oriented SVM method was76.83%, object-oriented ISODATA clustering was84.71%, object-oriented SVM method was88.44%). In addition, this proposed methodhas a good applicability for different sensors images.