Spectral Similarity Based Multiscale Spatial-Spectral Preprocessing Framework for Hyperspectral Image Classification

Abstract

Hyperspectral imaging represents an advanced technology that offers an extensive array of spectral data concerning various materials. Each pixel within a hyperspectral image encompasses reflectance or transmittance values spanning a spectrum of wavelengths, thereby constructing a spectral signature or spectral curve. Despite the high spectral resolution inherent in hyperspectral images, their spatial resolution frequently remains limited, resulting in a mixture of spectral information within the spectral signatures. This situation presents a significant obstacle to achieving precise hyperspectral image classification, given that both spectral and spatial information play pivotal roles in this endeavor. In the present investigation, a novel spectral-spatial preprocessing strategy is introduced, employing a multiscale filtering technique based on spectral similarity to enhance the accuracy of hyperspectral image classification. The methodology entails performing a neighborhood operation for each target pixel vector, predicated on their spectral resemblance. This operation assigns higher priority to more similar pixels within the neighborhood window to establish the new spectral curve of the pixel of interest. The resultant spectral curves effectively amalgamate both spatial and spectral information and are subsequently utilized during the classification process instead of the original spectral curves. The study incorporates established spectral similarity metrics alongside an innovative metric grounded in Fr & eacute;chet distance to calculate spectral similarities. The outcomes derived from these metrics are juxtaposed to assess their efficacy in ameliorating the accuracy of hyperspectral image classification. Moreover, the classification performance is evaluated utilizing kernel extreme learning machine and support vector classifiers across four distinct hyperspectral image datasets. The findings underscore that, particularly when confronted with constraints related to small sample sizes, the proposed spectral-spatial preprocessing technique markedly enhances the classification accuracy of hyperspectral images.