Using Savitsky-Golay Smoothing Filter in Hyperspectral Data Compression by Curve Fitting

Hyperspectral images (HS) are collected images of earth's surface over hundreds of narrow and close together spectral bands. This type of image should to be compressed because of the high between bands correlation and transmission of a very high amount of storage. There are different methods for compressing in the spatial or spectrum space that can be lossy or lossless. But it should be considered that in the field of remote sensing spectral data in hyperspectral images is more important than spatial data, so the compression should be performed somehow the spectral information of these images is well preserved. Our proposed method in this paper is a lossy compression technique that is based on the use of the curve fitting. It is recognized that the compression method using curve fitting has very good performance compared to other methods such as the principal component analysis (PCA). In this method, the spectral signature of each pixel from the original data is smoothed by using Savitzky-Golay smoothing filter with the most appropriate window size and smoothing polynomial grade, then a rational curve with the best degrees of numerator and denominator polynomials. Coefficients of the numerator and denominator of this function are considered as new features and thus the original data is compressed well. The results indicate that compressed data after recovery has a very close resemblance to the original data.

I. Introduction

The Spectrum of each material considered as a unique feature like fingerprints, which is why the graph of the brightness of each substance listed in spectral range, called spectral signature or spectral reflectance curve (SRC). Hyper spectral images contains very important spectral data since they are collecting in a numerous spectra's, also this data helps the clarity of these images. On the other hand, since data collected in narrow and close together spectra's, there are a lot of correlation between bands of these images, hence for easy storage and transfer, images are better to be compressed first. It should be noted that remove redundancies should be carried out somehow spectral information well preserved [1]. Compression could be lossless or lossy [2], [3].