A bag-of-visual-words model based framework for object-oriented land-cover classification

Abstract

Land-cover composition and change are important factors that affect global ecosystem. As an effective means for Earth observation, remote sensing technique has been widely applied in extracting land-cover information and in monitoring land-use and land-cover change, among which image classification becomes a key issue. Most existing studies about object-oriented classification use traditional low-level feature extraction methods or statistics of low-level features to represent objects in an image, which, to a large extent, loses the information in remote sensing images. Therefore, in order to facilitate better description of these objects in object-oriented classification, this paper introduces a state-of-theart feature representation method called bag-of-visual-words (BOVW) to construct the middle-level representations instead of low-level features. Based on the idea of BOVW, this paper proposes a BOVW based framework for objectoriented land-cover classification. For a given remote sensing image, it first applies a pixel-level local feature extraction strategy to construct a visual vocabulary by K-means clustering with each cluster as a visual word. Then the image is segmented into objects and each object is represented as a histogram of visual word occurrences by mapping the local pixel-level features in this object to the learned visual words. Finally, the calculated histogram is considered as the final representation of an object which can be used for further classification tasks. Experimental results on a SPOT5 satellite image, acquired from the Changping County in Beijing, China, in 2002, show that the proposed method is superior to the traditional low-level feature based method in classification accuracy by about 2%.