Spatial residual clustering and entropy based ranking for hyperspectral band selection
Though the Hyper-spectral images (HSI) are associated with rich spectral information for discriminating the class-specific objects, the high dimensional data generates Hughes effect for additional processing. So, during pre-processing, band Selection (BS) is done to lower the dimension. The proposed...
Main Authors: | , , |
---|---|
Format: | Article |
Language: | English |
Published: |
Taylor & Francis Group
2020-06-01
|
Series: | European Journal of Remote Sensing |
Subjects: | |
Online Access: | http://dx.doi.org/10.1080/22797254.2019.1703559 |
id |
doaj-480e11c1562f49f19f374ba66935df2e |
---|---|
record_format |
Article |
spelling |
doaj-480e11c1562f49f19f374ba66935df2e2020-11-25T03:47:02ZengTaylor & Francis GroupEuropean Journal of Remote Sensing2279-72542020-06-0153S1829210.1080/22797254.2019.17035591703559Spatial residual clustering and entropy based ranking for hyperspectral band selectionKishore Raju K.0Saradhi Varma G. P.1Rajya Lakshmi D.2JNTUKS.R.K.R Engineering CollegeJNTUK UCENThough the Hyper-spectral images (HSI) are associated with rich spectral information for discriminating the class-specific objects, the high dimensional data generates Hughes effect for additional processing. So, during pre-processing, band Selection (BS) is done to lower the dimension. The proposed unsupervised BS technique follows ‘3ʹ subsequent steps, such as 1) spatial residual (Gaussian) filtering, 2) spatial feature-based band clustering (using K-means) and 3) Entropy-based ranking. Initially, the bands are filtered and the self residual images are obtained. The filtered residual images are mapped to k clusters, where each cluster represents unique spatial information. Further, each cluster undergoes selection through the information-theoretic (Entropy) approach. Also, from the selected optimal bands, SpatioSpectral information is extracted to appraise the performance using Support Vector Machine (SVM) classifiers with different state-of-art approaches. Performance measures like Average Accuracy (AA), Kappa (κ) and Overall Accuracy (OA) are contrasted and are presented in tables. The experiment shows propitious results as contrasted to other approaches, and suggests that pre clustering of the bands is informative as the adjacent bands are strongly correlated. Also, on comparing other approaches, the proposed one is computationally cheap and faster, which may best suit for online (in-camera) BS purposes.http://dx.doi.org/10.1080/22797254.2019.1703559clusteringdimensionality reductionentropy-based rankinghyperspectral imageresidual filteringunsupervised band selection |
collection |
DOAJ |
language |
English |
format |
Article |
sources |
DOAJ |
author |
Kishore Raju K. Saradhi Varma G. P. Rajya Lakshmi D. |
spellingShingle |
Kishore Raju K. Saradhi Varma G. P. Rajya Lakshmi D. Spatial residual clustering and entropy based ranking for hyperspectral band selection European Journal of Remote Sensing clustering dimensionality reduction entropy-based ranking hyperspectral image residual filtering unsupervised band selection |
author_facet |
Kishore Raju K. Saradhi Varma G. P. Rajya Lakshmi D. |
author_sort |
Kishore Raju K. |
title |
Spatial residual clustering and entropy based ranking for hyperspectral band selection |
title_short |
Spatial residual clustering and entropy based ranking for hyperspectral band selection |
title_full |
Spatial residual clustering and entropy based ranking for hyperspectral band selection |
title_fullStr |
Spatial residual clustering and entropy based ranking for hyperspectral band selection |
title_full_unstemmed |
Spatial residual clustering and entropy based ranking for hyperspectral band selection |
title_sort |
spatial residual clustering and entropy based ranking for hyperspectral band selection |
publisher |
Taylor & Francis Group |
series |
European Journal of Remote Sensing |
issn |
2279-7254 |
publishDate |
2020-06-01 |
description |
Though the Hyper-spectral images (HSI) are associated with rich spectral information for discriminating the class-specific objects, the high dimensional data generates Hughes effect for additional processing. So, during pre-processing, band Selection (BS) is done to lower the dimension. The proposed unsupervised BS technique follows ‘3ʹ subsequent steps, such as 1) spatial residual (Gaussian) filtering, 2) spatial feature-based band clustering (using K-means) and 3) Entropy-based ranking. Initially, the bands are filtered and the self residual images are obtained. The filtered residual images are mapped to k clusters, where each cluster represents unique spatial information. Further, each cluster undergoes selection through the information-theoretic (Entropy) approach. Also, from the selected optimal bands, SpatioSpectral information is extracted to appraise the performance using Support Vector Machine (SVM) classifiers with different state-of-art approaches. Performance measures like Average Accuracy (AA), Kappa (κ) and Overall Accuracy (OA) are contrasted and are presented in tables. The experiment shows propitious results as contrasted to other approaches, and suggests that pre clustering of the bands is informative as the adjacent bands are strongly correlated. Also, on comparing other approaches, the proposed one is computationally cheap and faster, which may best suit for online (in-camera) BS purposes. |
topic |
clustering dimensionality reduction entropy-based ranking hyperspectral image residual filtering unsupervised band selection |
url |
http://dx.doi.org/10.1080/22797254.2019.1703559 |
work_keys_str_mv |
AT kishorerajuk spatialresidualclusteringandentropybasedrankingforhyperspectralbandselection AT saradhivarmagp spatialresidualclusteringandentropybasedrankingforhyperspectralbandselection AT rajyalakshmid spatialresidualclusteringandentropybasedrankingforhyperspectralbandselection |
_version_ |
1724503809964638208 |