Lossless predictive compression of hyperspectral images

Hongqiang Wang, Khalid Sayood

Research output: Chapter in Book/Report/Conference proceedingChapter

7 Citations (Scopus)

Abstract

After almost three decades of successful data acquisition using multispectral sensors the first space based hyperspectral sensors were launched in 2000 on the NASA EO-1 satellite. However, airborne hyperspectral sensors such as AVIRIS, among others, have been generating useful data for many years. The advent of the space-borne ALI and Hyperion sensors as well as the successes of AVIRIS presage the development of many more hyperspectral instruments. Furthermore the success of multispectral imagers such as the Enhanced Thematic Mapper Plus (EMT+) on the LANDSAT-7 mission and the modestly named 36 band MODIS (Moderate Resolution Imaging Spectroradiometer) instrument aboard the Terra satellite and the Aqua spacecraft promises the deployment of significant numbers of other such instruments. The use of multispectral and hyperspectral sensors, while opening the door to multiple applications in climate observation, environmental monitoring, and resource mapping, among others, also means the generation of huge amounts of data that needs to be accommodated by transmission and distribution facilities that cannot economically handle this level of data. This means that compression, always a pressing concern [1], is now imperative. While in many cases the use of lossy compression may be unavoidable, it is important that the design always include the possibility of lossless recovery. Much effort usually has gone into the reduction of noise in the instruments. The voluntary addition of noise due to compression can be a bitter pill to swallow. Compression is needed, and can be applied, in several different places from where the images are acquired to the end-user. At the point of acquisition compression may be required under several different scenarios. The satellite carrying the sensor may not be in continuous contact with ground stations. In the interval between contacts the data has to stored on board. If these intervals are of any significant length the amount of data generated is likely to be very high and compression becomes imperative. Any compression at this point has to be lossless. Even if there is relatively frequent contact between the satellite and the ground station if the portion of the bandwidth available to the instrument is less than the raw data rate again compression is required, and again this compression has to be lossless. Once the data is on the ground it needs to be archived. This is another point at which lossless compression may be needed. Finally, the data has to be distributed to end-users. Depending on the amount of data desired by a particular end-user compression may or may not be required. Furthermore, depending on the application of interest to the enduser this compression can be lossy or lossless. Thus, while it can be argued for certain applications that given a stringent enough distortion constraint lossy compression may be possible there are many scenarios in which only lossless compression will be acceptable. Various lossless compression schemes based on transform coding, vector quantization [2, 3], and predictive coding have been proposed. In this chapter we examine a number of lossless compression schemes based on predictive coding. Most lossless compression schemes include prediction as part of the algorithm. Strong local correlations in the image allow prediction of the pixel being encoded. The difference between the pixel and the prediction, known as the prediction error or prediction residual, usually has a much lower first order entropy than the pixel itself and therefore, can be encoded with fewer bits. Well known predictive lossless image compression techniques developed for use with natural images include CALIC {Context-based Adaptive Lossless Image Compression) [4, 5] and LOCO-I {Low Complexity Lossless Compression for Images) [6] which is part of JPEG-LS, the ISO/ITU standard for lossless and near-lossless compression of images. These, as well as others, have their three dimensional counterparts. In recent years, various lossless hyperspectral image compression schemes using reversible integer wavelet transforms have been proposed. These include EZW {Embedded Zerotree Wavelet) [7] and SPIHT {Set Partitioning in Hierarchical Trees) compression [8]. These schemes takes advantage of the fact that when an image is decomposed based on its frequency the coefficients at higher frequency are closely related to the coefficients at lower frequency. Wavelet decomposition and predictive coding can be combined in two ways. The prediction error residuals can be encoded using a wavelet coder [9], or the decomposition of the images can be used as a preprocessing step prior to prediction. In the following we describe some of the more popular predictive coding schemes as they have been applied to hyperspectral compression.

Original languageEnglish (US)
Title of host publicationHyperspectral Data Compression
PublisherSpringer US
Pages35-55
Number of pages21
ISBN (Print)0387285792, 9780387285795
DOIs
StatePublished - Dec 1 2006

Fingerprint

Sensors
Image compression
Satellites
Pixels
Wavelet decomposition
Vector quantization
Image sensors
Wavelet transforms
NASA
Spacecraft
Data acquisition
Entropy
Bandwidth
Imaging techniques
Recovery
Monitoring

ASJC Scopus subject areas

  • Computer Science(all)

Cite this

Wang, H., & Sayood, K. (2006). Lossless predictive compression of hyperspectral images. In Hyperspectral Data Compression (pp. 35-55). Springer US. https://doi.org/10.1007/0-387-28600-4_2

Lossless predictive compression of hyperspectral images. / Wang, Hongqiang; Sayood, Khalid.

Hyperspectral Data Compression. Springer US, 2006. p. 35-55.

Research output: Chapter in Book/Report/Conference proceedingChapter

Wang, H & Sayood, K 2006, Lossless predictive compression of hyperspectral images. in Hyperspectral Data Compression. Springer US, pp. 35-55. https://doi.org/10.1007/0-387-28600-4_2
Wang H, Sayood K. Lossless predictive compression of hyperspectral images. In Hyperspectral Data Compression. Springer US. 2006. p. 35-55 https://doi.org/10.1007/0-387-28600-4_2
Wang, Hongqiang ; Sayood, Khalid. / Lossless predictive compression of hyperspectral images. Hyperspectral Data Compression. Springer US, 2006. pp. 35-55
@inbook{8b375b4e6a22446da12c1b9a5cc2095c,
title = "Lossless predictive compression of hyperspectral images",
abstract = "After almost three decades of successful data acquisition using multispectral sensors the first space based hyperspectral sensors were launched in 2000 on the NASA EO-1 satellite. However, airborne hyperspectral sensors such as AVIRIS, among others, have been generating useful data for many years. The advent of the space-borne ALI and Hyperion sensors as well as the successes of AVIRIS presage the development of many more hyperspectral instruments. Furthermore the success of multispectral imagers such as the Enhanced Thematic Mapper Plus (EMT+) on the LANDSAT-7 mission and the modestly named 36 band MODIS (Moderate Resolution Imaging Spectroradiometer) instrument aboard the Terra satellite and the Aqua spacecraft promises the deployment of significant numbers of other such instruments. The use of multispectral and hyperspectral sensors, while opening the door to multiple applications in climate observation, environmental monitoring, and resource mapping, among others, also means the generation of huge amounts of data that needs to be accommodated by transmission and distribution facilities that cannot economically handle this level of data. This means that compression, always a pressing concern [1], is now imperative. While in many cases the use of lossy compression may be unavoidable, it is important that the design always include the possibility of lossless recovery. Much effort usually has gone into the reduction of noise in the instruments. The voluntary addition of noise due to compression can be a bitter pill to swallow. Compression is needed, and can be applied, in several different places from where the images are acquired to the end-user. At the point of acquisition compression may be required under several different scenarios. The satellite carrying the sensor may not be in continuous contact with ground stations. In the interval between contacts the data has to stored on board. If these intervals are of any significant length the amount of data generated is likely to be very high and compression becomes imperative. Any compression at this point has to be lossless. Even if there is relatively frequent contact between the satellite and the ground station if the portion of the bandwidth available to the instrument is less than the raw data rate again compression is required, and again this compression has to be lossless. Once the data is on the ground it needs to be archived. This is another point at which lossless compression may be needed. Finally, the data has to be distributed to end-users. Depending on the amount of data desired by a particular end-user compression may or may not be required. Furthermore, depending on the application of interest to the enduser this compression can be lossy or lossless. Thus, while it can be argued for certain applications that given a stringent enough distortion constraint lossy compression may be possible there are many scenarios in which only lossless compression will be acceptable. Various lossless compression schemes based on transform coding, vector quantization [2, 3], and predictive coding have been proposed. In this chapter we examine a number of lossless compression schemes based on predictive coding. Most lossless compression schemes include prediction as part of the algorithm. Strong local correlations in the image allow prediction of the pixel being encoded. The difference between the pixel and the prediction, known as the prediction error or prediction residual, usually has a much lower first order entropy than the pixel itself and therefore, can be encoded with fewer bits. Well known predictive lossless image compression techniques developed for use with natural images include CALIC {Context-based Adaptive Lossless Image Compression) [4, 5] and LOCO-I {Low Complexity Lossless Compression for Images) [6] which is part of JPEG-LS, the ISO/ITU standard for lossless and near-lossless compression of images. These, as well as others, have their three dimensional counterparts. In recent years, various lossless hyperspectral image compression schemes using reversible integer wavelet transforms have been proposed. These include EZW {Embedded Zerotree Wavelet) [7] and SPIHT {Set Partitioning in Hierarchical Trees) compression [8]. These schemes takes advantage of the fact that when an image is decomposed based on its frequency the coefficients at higher frequency are closely related to the coefficients at lower frequency. Wavelet decomposition and predictive coding can be combined in two ways. The prediction error residuals can be encoded using a wavelet coder [9], or the decomposition of the images can be used as a preprocessing step prior to prediction. In the following we describe some of the more popular predictive coding schemes as they have been applied to hyperspectral compression.",
author = "Hongqiang Wang and Khalid Sayood",
year = "2006",
month = "12",
day = "1",
doi = "10.1007/0-387-28600-4_2",
language = "English (US)",
isbn = "0387285792",
pages = "35--55",
booktitle = "Hyperspectral Data Compression",
publisher = "Springer US",

}

TY - CHAP

T1 - Lossless predictive compression of hyperspectral images

AU - Wang, Hongqiang

AU - Sayood, Khalid

PY - 2006/12/1

Y1 - 2006/12/1

N2 - After almost three decades of successful data acquisition using multispectral sensors the first space based hyperspectral sensors were launched in 2000 on the NASA EO-1 satellite. However, airborne hyperspectral sensors such as AVIRIS, among others, have been generating useful data for many years. The advent of the space-borne ALI and Hyperion sensors as well as the successes of AVIRIS presage the development of many more hyperspectral instruments. Furthermore the success of multispectral imagers such as the Enhanced Thematic Mapper Plus (EMT+) on the LANDSAT-7 mission and the modestly named 36 band MODIS (Moderate Resolution Imaging Spectroradiometer) instrument aboard the Terra satellite and the Aqua spacecraft promises the deployment of significant numbers of other such instruments. The use of multispectral and hyperspectral sensors, while opening the door to multiple applications in climate observation, environmental monitoring, and resource mapping, among others, also means the generation of huge amounts of data that needs to be accommodated by transmission and distribution facilities that cannot economically handle this level of data. This means that compression, always a pressing concern [1], is now imperative. While in many cases the use of lossy compression may be unavoidable, it is important that the design always include the possibility of lossless recovery. Much effort usually has gone into the reduction of noise in the instruments. The voluntary addition of noise due to compression can be a bitter pill to swallow. Compression is needed, and can be applied, in several different places from where the images are acquired to the end-user. At the point of acquisition compression may be required under several different scenarios. The satellite carrying the sensor may not be in continuous contact with ground stations. In the interval between contacts the data has to stored on board. If these intervals are of any significant length the amount of data generated is likely to be very high and compression becomes imperative. Any compression at this point has to be lossless. Even if there is relatively frequent contact between the satellite and the ground station if the portion of the bandwidth available to the instrument is less than the raw data rate again compression is required, and again this compression has to be lossless. Once the data is on the ground it needs to be archived. This is another point at which lossless compression may be needed. Finally, the data has to be distributed to end-users. Depending on the amount of data desired by a particular end-user compression may or may not be required. Furthermore, depending on the application of interest to the enduser this compression can be lossy or lossless. Thus, while it can be argued for certain applications that given a stringent enough distortion constraint lossy compression may be possible there are many scenarios in which only lossless compression will be acceptable. Various lossless compression schemes based on transform coding, vector quantization [2, 3], and predictive coding have been proposed. In this chapter we examine a number of lossless compression schemes based on predictive coding. Most lossless compression schemes include prediction as part of the algorithm. Strong local correlations in the image allow prediction of the pixel being encoded. The difference between the pixel and the prediction, known as the prediction error or prediction residual, usually has a much lower first order entropy than the pixel itself and therefore, can be encoded with fewer bits. Well known predictive lossless image compression techniques developed for use with natural images include CALIC {Context-based Adaptive Lossless Image Compression) [4, 5] and LOCO-I {Low Complexity Lossless Compression for Images) [6] which is part of JPEG-LS, the ISO/ITU standard for lossless and near-lossless compression of images. These, as well as others, have their three dimensional counterparts. In recent years, various lossless hyperspectral image compression schemes using reversible integer wavelet transforms have been proposed. These include EZW {Embedded Zerotree Wavelet) [7] and SPIHT {Set Partitioning in Hierarchical Trees) compression [8]. These schemes takes advantage of the fact that when an image is decomposed based on its frequency the coefficients at higher frequency are closely related to the coefficients at lower frequency. Wavelet decomposition and predictive coding can be combined in two ways. The prediction error residuals can be encoded using a wavelet coder [9], or the decomposition of the images can be used as a preprocessing step prior to prediction. In the following we describe some of the more popular predictive coding schemes as they have been applied to hyperspectral compression.

AB - After almost three decades of successful data acquisition using multispectral sensors the first space based hyperspectral sensors were launched in 2000 on the NASA EO-1 satellite. However, airborne hyperspectral sensors such as AVIRIS, among others, have been generating useful data for many years. The advent of the space-borne ALI and Hyperion sensors as well as the successes of AVIRIS presage the development of many more hyperspectral instruments. Furthermore the success of multispectral imagers such as the Enhanced Thematic Mapper Plus (EMT+) on the LANDSAT-7 mission and the modestly named 36 band MODIS (Moderate Resolution Imaging Spectroradiometer) instrument aboard the Terra satellite and the Aqua spacecraft promises the deployment of significant numbers of other such instruments. The use of multispectral and hyperspectral sensors, while opening the door to multiple applications in climate observation, environmental monitoring, and resource mapping, among others, also means the generation of huge amounts of data that needs to be accommodated by transmission and distribution facilities that cannot economically handle this level of data. This means that compression, always a pressing concern [1], is now imperative. While in many cases the use of lossy compression may be unavoidable, it is important that the design always include the possibility of lossless recovery. Much effort usually has gone into the reduction of noise in the instruments. The voluntary addition of noise due to compression can be a bitter pill to swallow. Compression is needed, and can be applied, in several different places from where the images are acquired to the end-user. At the point of acquisition compression may be required under several different scenarios. The satellite carrying the sensor may not be in continuous contact with ground stations. In the interval between contacts the data has to stored on board. If these intervals are of any significant length the amount of data generated is likely to be very high and compression becomes imperative. Any compression at this point has to be lossless. Even if there is relatively frequent contact between the satellite and the ground station if the portion of the bandwidth available to the instrument is less than the raw data rate again compression is required, and again this compression has to be lossless. Once the data is on the ground it needs to be archived. This is another point at which lossless compression may be needed. Finally, the data has to be distributed to end-users. Depending on the amount of data desired by a particular end-user compression may or may not be required. Furthermore, depending on the application of interest to the enduser this compression can be lossy or lossless. Thus, while it can be argued for certain applications that given a stringent enough distortion constraint lossy compression may be possible there are many scenarios in which only lossless compression will be acceptable. Various lossless compression schemes based on transform coding, vector quantization [2, 3], and predictive coding have been proposed. In this chapter we examine a number of lossless compression schemes based on predictive coding. Most lossless compression schemes include prediction as part of the algorithm. Strong local correlations in the image allow prediction of the pixel being encoded. The difference between the pixel and the prediction, known as the prediction error or prediction residual, usually has a much lower first order entropy than the pixel itself and therefore, can be encoded with fewer bits. Well known predictive lossless image compression techniques developed for use with natural images include CALIC {Context-based Adaptive Lossless Image Compression) [4, 5] and LOCO-I {Low Complexity Lossless Compression for Images) [6] which is part of JPEG-LS, the ISO/ITU standard for lossless and near-lossless compression of images. These, as well as others, have their three dimensional counterparts. In recent years, various lossless hyperspectral image compression schemes using reversible integer wavelet transforms have been proposed. These include EZW {Embedded Zerotree Wavelet) [7] and SPIHT {Set Partitioning in Hierarchical Trees) compression [8]. These schemes takes advantage of the fact that when an image is decomposed based on its frequency the coefficients at higher frequency are closely related to the coefficients at lower frequency. Wavelet decomposition and predictive coding can be combined in two ways. The prediction error residuals can be encoded using a wavelet coder [9], or the decomposition of the images can be used as a preprocessing step prior to prediction. In the following we describe some of the more popular predictive coding schemes as they have been applied to hyperspectral compression.

UR - http://www.scopus.com/inward/record.url?scp=51849110694&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=51849110694&partnerID=8YFLogxK

U2 - 10.1007/0-387-28600-4_2

DO - 10.1007/0-387-28600-4_2

M3 - Chapter

SN - 0387285792

SN - 9780387285795

SP - 35

EP - 55

BT - Hyperspectral Data Compression

PB - Springer US

ER -