TY - CHAP
T1 - Constrained band selection for hyperspectral imaging
AU - Chang, Chein I.
N1 - Publisher Copyright:
© 2023 John Wiley & Sons, Inc.
PY - 2022/11/11
Y1 - 2022/11/11
N2 - Due to the use of hundreds of contiguous spectral bands provided by a hyperspectral imaging sensor, a hyperspectral image (HSI) generally contains wealthy spectral information for data processing. However, such very high inter-band correlation also comes with significant redundancy which results in enormous unnecessary data information. Band selection (BS) is one of major techniques to reduce high dimensionality of a HSI. This chapter is particularly devoted to two constrained band selection (CBS) techniques depending upon how constraints are imposed on BS. One is an application-independent BS, referred to as band-constrained BS (BCBS) which imposes constraints on the bands to be selected without specifying particular applications. Two versions of BCBS are developed, band vector-based BCBS, to be called constrained energy minimization (CEM)-based CBS (CEM-CBS), which uses CEM to constrain a band to be selected as a band vector and band image-based BCBS, to be called linearly constrained minimum variance-based constrained band selection (LCMV-CBS), which takes advantage of LCMV to constrain a band image to be selected as a matrix so as to reduce computational complexity. Since a HSI may have a very large size, representing an HSI as a band vector produces a very large vector dimensionality which results in excessive computing time. LCMV-CBS is developed to mitigate this dilemma. The other type of CBS is a target detection-specified BS, referred to as target-constrained BS (TCBS) which imposes constraints on targets to be detected. In analogy with BCBS, there are also two versions of TCBS that can be developed, a single target vector-based TCBS which uses CEM to constrain the desired target vector to be detected to select bands and a multiple target matrix-based TCBS which implements LCMV to constrain multiple target vectors to be detected to select bands. Similar treatments developed for BCBS can be also applied to TCBS with constrained band vectors and band images replaced by single target vector and multiple target matrix, respectively. In this chapter, both BCBS and TCBS will be studied in detail along with experiments conducted for demonstration.
AB - Due to the use of hundreds of contiguous spectral bands provided by a hyperspectral imaging sensor, a hyperspectral image (HSI) generally contains wealthy spectral information for data processing. However, such very high inter-band correlation also comes with significant redundancy which results in enormous unnecessary data information. Band selection (BS) is one of major techniques to reduce high dimensionality of a HSI. This chapter is particularly devoted to two constrained band selection (CBS) techniques depending upon how constraints are imposed on BS. One is an application-independent BS, referred to as band-constrained BS (BCBS) which imposes constraints on the bands to be selected without specifying particular applications. Two versions of BCBS are developed, band vector-based BCBS, to be called constrained energy minimization (CEM)-based CBS (CEM-CBS), which uses CEM to constrain a band to be selected as a band vector and band image-based BCBS, to be called linearly constrained minimum variance-based constrained band selection (LCMV-CBS), which takes advantage of LCMV to constrain a band image to be selected as a matrix so as to reduce computational complexity. Since a HSI may have a very large size, representing an HSI as a band vector produces a very large vector dimensionality which results in excessive computing time. LCMV-CBS is developed to mitigate this dilemma. The other type of CBS is a target detection-specified BS, referred to as target-constrained BS (TCBS) which imposes constraints on targets to be detected. In analogy with BCBS, there are also two versions of TCBS that can be developed, a single target vector-based TCBS which uses CEM to constrain the desired target vector to be detected to select bands and a multiple target matrix-based TCBS which implements LCMV to constrain multiple target vectors to be detected to select bands. Similar treatments developed for BCBS can be also applied to TCBS with constrained band vectors and band images replaced by single target vector and multiple target matrix, respectively. In this chapter, both BCBS and TCBS will be studied in detail along with experiments conducted for demonstration.
UR - http://www.scopus.com/inward/record.url?scp=85147789220&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85147789220&partnerID=8YFLogxK
U2 - 10.1002/9781119687788.ch4
DO - 10.1002/9781119687788.ch4
M3 - Chapter
AN - SCOPUS:85147789220
SN - 9781119687788
SP - 109
EP - 146
BT - Advances in Hyperspectral Image Processing Techniques
PB - Wiley-Blackwell
ER -