Band Selection-Based Dimensionality Reduction for Change Detection in Multi-Temporal Hyperspectral Images

Sicong Liu,Alim Samat,Qian Du,Haiyan Pan,Xiaolong Ma,Xiaohua Tong

doi:10.3390/rs9101008

Abstract

This paper proposes to use band selection-based dimensionality reduction (BS-DR) technique in addressing a challenging multi-temporal hyperspectral images change detection (HSI-CD) problem. The aim of this work is to analyze and evaluate in detail the CD performance by selecting the most informative band subset from the original high-dimensional data space. In particular, for cases where ground reference data are available or unavailable, either supervised or unsupervised CD approaches are designed. The following sub-problems in HSI-CD are investigated, including: (1) the estimated number of multi-class changes; (2) the binary CD; (3) the multiple CD; (4) the estimated optimal number of selected bands; and (5) computational efficiency. The main contribution of this paper is to provide for the first time a thorough analysis of the impacts of band selection on the HSI-CD problem, thus to fix the gap in the state-of-the-art techniques either by simply utilizing the full dimensionality of the data or exploring a complex hierarchical change analysis. It is applicable to CD problems in multispectral or PolSAR images when the feature space is expanded for discriminant feature extraction. Two real multi-temporal hyperspectral Hyperion datasets are used to validate the proposed approaches. Quantitative and qualitative experimental results demonstrated that by selecting a subset of the most informative and distinct spectral bands, the proposed approaches offered better CD performance than the state-of-the-art techniques using original full bands, without losing the change representative and discriminable capabilities of a detector.

Highlights

Next-generation hyperspectral sensors onboard airborne and spaceborne crafts can acquire hyperspectral images (HSIs) through dense spectral sampling (e.g., 1–10 nm) over a wide wavelength spectrum (e.g., 400–2500 nm) [1]
We address for the first time the challenging hyperspectral images change detection (HSI-change detection (CD)) problem in a low-dimensional feature space by using band selection-based Dimensionality reduction (DR) (BS-DR) algorithms
The unsupervised LP algorithm and the supervised minimum estimated abundance covariance (MEAC) algorithm were applied to XD, while3v.2a.rRyeisnuglttshoenstehleeYcatendchbenagndWset(lia.en.d, MAg)rficruolmtur1altoD3at0a.seTthe first 20 selected bands in the two algorithms are highliTghheteudnsuinpebrlvuiseed(uLnPsaulpgoerrivthismeda)ndanthde sruepder(vsuispederMviEsAedC)ailngoFriitghumrewe5rbe.apWpleiedcatno XoDb,swehrvile that they avraerylioncgattheedseinlecdteifdfebraenndtsh(ii.ge.h, lMy)cforormrel1attoed30s. pTehcetfriarsltr2e0gsieolnecste(di.eb.a, nSd1s–Sin5)t.heTthwios adlegmoriothnmstsraatres the effecthivigehnleigshstoedf tihnebaludeo(putnesdupbearnvdisesde)leacntdiornedap(spurpoearcvhisetod)eixntrFaigcut rtehe5bm

Summary

Introduction

Next-generation hyperspectral sensors onboard airborne and spaceborne crafts can acquire hyperspectral images (HSIs) through dense spectral sampling (e.g., 1–10 nm) over a wide wavelength spectrum (e.g., 400–2500 nm) [1]. Different from the traditional multispectral images, which characterized coarse spectral resolution in several broad spectral channels (i.e., bands), the detailed spectral sampling in hyperspectral imaging results in hundreds or even thousands of contiguous spectral bands that dramatically increases data storage volume and the ensuing data processing complexity. “Hughes phenomenon”, i.e., with a fixed number of training samples, the predictive power of a classifier reduces as the dimensionality increases [2,3]. Such redundant information may affect how the user-interested information has been represented and detected to a great extent. The aforementioned challenges have raised many issues about handling hyperspectral data in different remote sensing tasks, e.g., classification, target detection, etc

Objectives

Methods

Results

Conclusion