Researches On Hyperspectral Image Classification Algorithms Based On Superpixel

With the rapid development of hyperspectral remote sensor technology, hyperspectral images(HSIs) have been widely used in different applications. Many researches have been done in HSI classification. The spectral information with hundreds of spectral wavelengths for each image pixel allows the classification of landcovers with improved accuracy. So, a lot of pixel-wise classifiers have been proposed using the spectral vectors for HSI. To improve the accuracy further, many spectral-spatial classification methods are investigated recently. How to utilize the spectral and spatial information is a non-trivial issue. In this paper, researches on how to extract spatial feature adaptively and use the spatial feature to improve the classification accuracy of HIS are described as follows:1. Spatial-spectral feature extracting method on superpixel: one superpixel can be regarded as a small region consisting of a number of pixels with similar spectral characteristics, balancing the conflicting goals of reducing image complexity through pixel grouping while avoiding under-segmentation. A superpixel representation greatly reduces the number of image primitives compared to the pixel representation. Moreover, superpixel segmentation provides the spatial support for computing region based features. This paper segments the hyperspectral image by superpixel algorithm and extracts the spatial features from each superpixel, which largely improved the classification accuracy.2. HSI classification using superpixel and extreme learning machines(ELMs): extreme learning machine is used for the classification of each mean feature to determine the class label of each superpixel. The framework can achieve very high efficiency and accuracy compared to other classification methods.3. Superpixel-based composite kernel for hyperspectral image classification: composite kernel methods can utilize both the spectral and the spatial information for the HSI classification. However, setting theoptimal spatial neighborhood for different spatial structures is a non-trivial issue. In order to adaptively exploit the spatialcontextual information, we utilize superpixel to obtain spatial information, the shape of each superpixel can be adaptivelychanged according to the spatial information. Superpixel can be regarded as a local neighborhood, whose size and shape canbe adaptively adjusted according to the spatial structures in the HSI. In this paper, we propose a superpixel-based compositekernel method, kernel functions constructed by two kernels and three kernels are implemented in the support vector machines, which can use composite kernel to effectively exploit the spectral and spatial information of the superpixel. Experiments on the real HSI demonstrate the outstanding performance of the proposed method.