Hyperspectral Image Classification Based on
Spectral-Spatial Feature Extraction
Zhen Ye, Lian Tan, Lin Bai
School of Electronics and Control Engineering
Chang'an University
Xi’an, China
yezhen525@126.com
Abstract—A novel hyperspectral classification algorithm
based on spectral-spatial feature extraction is proposed. First,
spectral-spatial features are extracted by Gabor transform in
PCA-projected space. Following that, Gabor-feature bands are
partitioned into multiple subsets. Afterwards, the adjacent
features in each subset are fused. Finally, the fused features are
processed by recursive filtering before feeding into support
vector machine (SVM) classifier. Experimental results
demonstrate that the proposed algorithm substantially
outperforms the traditional and state-of-the-art methods.
Keywords—hyperspectral image; classification; Gabor
feature extraction; image fusion; recursive filtering
I. INTRODUCTION
Hyperspectral imagery (HSI) records hundreds of spectral
bands for each pixel [1]. Due to the detailed spectral
information, HSI can be used to efficiently distinguish land-
cover types. However, the HSI often has a much greater
dimensionality than the number of available training
samples. The lack of training samples and the high
computational burden caused by high-dimensional data
processing are inevitable obstacles for designing HSI
classifiers. To avoid this “curse of dimensionality” problem,
dimensionality reduction is often employed to project the
HSI data onto a lower-dimensional feature space. Principal
component analysis (PCA) [2] is traditional methods used
for HSI dimensionality reduction. Since there is a high
probability that two adjacent pixels belong to the same HSI
class, spatial information can be helpful to create accurate
classification maps. Recently, two-dimensional Gabor
features extracted in PCA-projected subspaces have been
successfully used for spectral-spatial feature extraction [3].
Kang et al. [4] perform image fusion and recursive filtering
(IFRF) for feature extraction with the outstanding
performance in terms of classification accuracy and
computational efficiency. This paper presents a novel
hyperspectral classification algorithm based on spectral-
spatial feature extraction. As the first contribution, a Gabor
technique in PCA-projected subspace is studied to exploit
spectral and spatial information of HSI. The second
contribution is that Gabor feature bands is first introduced
into IFRF framework. Experimental results demonstrate that
the proposed framework is an innovative system for finer
features and efficient classification compared with the
traditional and other advanced spectral-spatial classification
methods.
II. APPROACH
As shown in Fig. 1, the proposed system consists of five
main steps: 1) extract spectral-spatial features by Gabor
transform in PCA-projected space; 2) partition Gabor-
feature bands into multiple subsets; 3) fuse the adjacent
features in each subset; 4) employ recursive filtering (RF)
on fused features; 5) employ support vector machine (SVM)
[5] for classification results.
Classification map
Gabor transform
Feature fusion
RF RF RF
Feature band
partitioning
SVM
Feature fusion Feature fusion
1
st
subset 2
st
subset 3
st
subset
Fig. 1 Flowchart of the proposed framework.
A. Gabor Transform
Gabor filter, which is a sinusoidal function modulated by
a Gaussian envelope, can be viewed as an orientation
dependent bandpass filter and can effectively capture the
orientation and scale of the physical structures for spatial
feature extraction [3]. However, the HSI contains a wealth
of spectral information over a wide range of the spectrum
and the wavelength interval are relatively small, which adds