A Novel Collaborative Representation Method for Hyperspectral Image Classification

With the rapid development of hyperspectral imaging technology, the effective utilization of abundant information inherent in hyperspectral images for achieving accurate classification has become a research focus. This paper proposes a hyperspectral image spatial-spectral fusion classification method based on collaborative representation. The proposed method calculates spatial information matrices and spectral information matrices between the hyperspectral image test samples and training samples separately. Subsequently, these matrices are fused to construct a collaborative representation-based hyperspectral image spatial-spectral fusion model, aiming to obtain collaborative representation coefficient matrices. These coefficient matrices not only reveal the spatial-spectral neighborhood structural characteristics of hyperspectral data more effectively but also facilitate the exploration of deep features. Finally, the test samples are reconstructed using the collaborative representation coefficients, and the classification results are obtained based on the minimal reconstruction error. In terms of method validation, a comparison is conducted with traditional SVM and SRC methods. Experimental results demonstrate the evident superiority of the proposed method in Hyperspectral Image classification performance.

I. Introduction

Hyperspectral remote sensing technology refers to employ imaging spectrometers to capture hyperspectral images composed of tens to hundreds of spectral bands. These captured images are then subjected to analysis and processing, enabling a detailed recognition of different features on the Earth's surface. A hyperspectral image is a three-dimensional representation comprising one spectral dimension and two spatial dimensions [1]. The spectral dimension encompasses spectral bands ranging from visible light to shortwave infrared, often numbering in the hundreds [2].