Joint kernelized sparse representation classification for hyperspectral imagery

Sun, He and Ren, Jinchang and Yan, Yijun and Zabalza, Jaime and Marshall, Stephen (2018) Joint kernelized sparse representation classification for hyperspectral imagery. In: Hyperspectral Imaging Applications (HSI) 2018, 2018-10-10 - 2018-10-11.

[thumbnail of Sun-etal-HSI2018-Joint-kernelized-sparse-representation-classification-for-hyperspectral-imagery]
Preview
Text. Filename: Sun_etal_HSI2018_Joint_kernelized_sparse_representation_classification_for_hyperspectral_imagery.pdf
Accepted Author Manuscript

Download (382kB)| Preview

Abstract

In recent years, the hyperspectral image (HSI) classification has received much attention due to its importance on the military applications, food quality assessment [1], and land cover analysis [2-5], etc. Multiple classifiers have been adopted to label pixels of HSI images, including support vector machine (SVM), random forest (RF), and recently, the deep learning methods. Considering that HSI pixels belonging to the same class are usually lying in a low-dimensional space, those pixels can be represented by training samples from the same class. Based on that, Sparse Representation Classification (SRC) methods have also introduced in the HSI imagery. For an unlabeled pixel, a few atoms from the constructed training dictionary can sparsely represent it. With the recovered sparse coefficients, the class label can be determined by the residual between the test pixel and its approximation. With the development of SRC in HIS [5, 6], there is one severe problem during the process of classification. Due to the high dimensions of the HSI data, it may result the Hughes phenomenon. Sufficient training samples are required to overcome the curse of dimensionality. However, sufficient training data are not always available in real application. For example, the ground truth labelling work for remote sensing data is rather inconvenient. Therefore, to solve the above problem, we decide to combine multiple types of features extracted from HSI data, and a joint kernelized SRC will be operated on those extracted features. The aim of our work is to improve the performance of SRC with less training samples.