Hyperspectral Image Classification Via Sparse Representation

With the continuous advancement of remote sensing technology and imaging equipment,hyperspectral images are widely used.Classification is a main branch of hyperspectral image processing and application.The goal is to assign unique pixel identifications to each pixel in the image.This article will focus on the hyperspectral images classification based on sparse representation and propose an effective classification algorithm.In order to overcome the problems caused by the traditional method of constructing a dictionary by stacking training samples,this paper proposes a dictionary learning algorithm ALSS based on adaptive local structure similarity,aiming at training a dictionary with high robustness.The algorithm not only considers the spatial relationship between pixels,but also improves the way of using the predefined and fixed Laplacian operator to describe the local spatial relationships among the pixels in the traditional method.It adaptively modify the local spatial relationships between pixels in the process of the dictionary learning.The dictionary learning algorithm,ALSS,has the following two advantages: taking into account the similarity relationship of the sparse coefficients itself;optimizing the local spatial relations in the dictionary learning process can make the dictionary more discriminative and more accurate in sparse representation.In addition,ALSS also divides the hyperspectral image into spatially adjacent but non-overlapping spatial blocks.Since each spatial block is usually composed of the same substance,the intra-block pixels can be linearly represented by a small number of common elements from the dictionary.Thus,this way not only combines spatial correlation information between hyperspectral image pixels but also provides conditions for parallel processing.By analyzing the classification results of the ALSS algorithm,it can be found that it's classification boundary is slightly jagged.Therefore,using regular window to segment the hyperspectral image data is not practical.Superpixels can include pixels whose positions are adjacent to each other and whose color,brightness,and texture features are similar in one block.These superpixels are uniform and do not overlap each other.This not only conforms to the actual conditions of hyperspectral image data,but also integrates The rich information,the most important thing is that it can perfectly combine with the ALSS algorithm to further improve the classification accuracy of the ALSS algorithm and make the classification result more smooth.Therefore,this paper proposes a hyperpixel-based hyperspectral image classification algorithm ALSS-SUP and ALSS-ERS.The difference between the two algorithms combined with superpixels: The first ALSS-SUP is change the spatial block in the ALSS algorithm model to the superpixels for processing;in the second ALSS-ERS,superpixel segmentation is used as a post-processing step.After the classification result of the ALSS algorithm is obtained,the classification result map is segmented as a number of superpixels,and then the majority voting method is used to update tag of superpixels and get the final classification result.In both methods,ALSS-SUP is higher than ALSS-ERS regardless of running time or classification accuracy,but both methods can improve the classification performance of ALSS.The proposed algorithm was used to classify the three commonly used hyperspectral image datasets.The experimental results verify their effectiveness.