Research On Multi-Level Classification Of High-Resolution Remote Sensing Image Based On Sharing Features

In recent years, high-resolution remote sensing images have been widely applied to many fields, such as agriculture, forestry, ocean, land resources, environmental monitoring and so on, so it is urgent for people to carry on research on classification of the high-resolution remote sensing images. Compared with medium and low resolution remote sensing images, high-resolution remote sensing images have rich shape, structure and texture information, so the conventional pixel-based classification method which mainly makes use of spectral information cannot meet the needs to process high-resolution remote sensing images, and the object-oriented classification method, developed during the last decade, has become the mainstream method. However, high resolution remote sensing images with abundant details generally have characteristics of great within-class differences and unobvious between-class differences. This brings challenge for their multiclass classification with high accuracy.The traditional multiclass classification algorithms, generally put samples of all classes into the multiclass classifier at the same time for training and then make use of the trained classifier to directly classify the test samples of all classes. This method treats all classes equally, and ignores the correlation among samples of different classes and the peculiarity of every single class, restricting the increase of classification accuracy to some extent. We notice that, usually when the interpretation experts visually interpret the remote sensing images, they don’t directly identify all classes of image objects, but firstly find out the most easily distinguishing class of samples, then find out the secondly easily distinguishing class of samples, until identify all classes of samples. On the other hand, image objects of different classes may have some similar characteristics which are called "sharing features" and can be used as the important information to distinguish between these classes sharing features and the other class.Simulating the visual interpretation, this paper proposes a multi-stage binary tree-structured classification algorithm based on sharing features. The multiclass classification problem is divided into multiple binary classification problems, by means of the GentleBoost algorithm sharing features are extracted to interpret objects of only one class at each binary classification stage, and each interpreted class will not participate in later classification. The proposed method makes use of the ’phase-out’ mechanism to complete the whole interpretation of a remote sensing image. Experimental results show that the multi-stage binary tree-structured classification algorithm can well simulate this visual interpretation process, extracting sharing features is helpful to improve the classification accuracy, and the multi-stage binary tree-structured classification algorithm based on sharing features has good classification result about high-resolution remote sensing images.