Espectro-Radiometría y Espectro-Radiometría y
Centro de Ciencias Humanas y SocialesConsejo Superior de Investigaciones CientíficasMadrid, 3-4 de Diciembre de 2009
Antonio PlazaAntonio PlazaUniversidad de ExtremaduraUniversidad de Extremadura
EE--mail: [email protected] mail: [email protected] http://www.umbc.edu/rssipl/people/aplazahttp://www.umbc.edu/rssipl/people/aplaza
Espectro-Radiometría y Teledetección Hiperespectral
Espectro-Radiometría y Teledetección Hiperespectral
ContentsContents1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
Hyperspectral image analysis: contents and distribution
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009
What is hyperspectral remote sensing?What is hyperspectral remote sensing?“Imaging Spectroscopy (hyperspectral remote sensing) deals with data from instruments acquiring reflectance images in a large (>40) number of narrow (<0.01 to 0.02 µm. in width), contiguous (i.e., adjacent) spectral bands allowing to derive the mineralogy of objects or obtaining information on soil, water and biochemical composition.”
“Imaging Spectroscopy (hyperspectral remote sensing) deals with data from instruments acquiring reflectance images in a large (>40) number of narrow (<0.01 to 0.02 µm. in width), contiguous (i.e., adjacent) spectral bands allowing to derive the mineralogy of objects or obtaining information on soil, water and biochemical composition.”
Introduction to hyperspectral imaging: increased spectral resolution
“The challenge of mapping minerals on any moon or planet with hyperspectral imaging relies onsub-pixel characterizationof spectral features fromthousands of possibilities.”
Lesson (not learned): The advantage of the concept (of imagingspectrometry) is that no morecommittees need to be formed to develop a rationale for particular spectral bands and to compromiseamongdisciplines whenit comes time to build the sensor’ [Goetz, 1992]
“The challenge of mapping minerals on any moon or planet with hyperspectral imaging relies onsub-pixel characterizationof spectral features fromthousands of possibilities.”
Lesson (not learned): The advantage of the concept (of imagingspectrometry) is that no morecommittees need to be formed to develop a rationale for particular spectral bands and to compromiseamongdisciplines whenit comes time to build the sensor’ [Goetz, 1992]
soil, water and biochemical composition.” Alexander F. H. Goetz, IGARSS 2006soil, water and biochemical composition.” Alexander F. H. Goetz, IGARSS 2006
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 1
Spectral mixture analysis: Determines the abundance of materials (e.g. precision agriculture).
Characterization: Determines variability of identified material (e.g. wet/dry sand, soil particle size effects).
Identification: Determines the unique identity of
Spectral mixture analysis: Determines the abundance of materials (e.g. precision agriculture).
Characterization: Determines variability of identified material (e.g. wet/dry sand, soil particle size effects).
Identification: Determines the unique identity of HyperspectralHyperspectral
Levels of Spectral Information in Remote Sensing
Ultraspectral
(1000’s of bands)
Ultraspectral
(1000’s of bands)
Introduction to hyperspectral imaging: increased spectral resolution
Identification: Determines the unique identity of the foregoing generic categories (e.g. land-cover or mineral mapping).
Discrimination: Determines generic categories of the foregoing classes.
Classification: Separates materials into spectrally similar groups (e.g., urban data classification).
Detection: Determines the presence of materials, objects, activities, or events.
Identification: Determines the unique identity of the foregoing generic categories (e.g. land-cover or mineral mapping).
Discrimination: Determines generic categories of the foregoing classes.
Classification: Separates materials into spectrally similar groups (e.g., urban data classification).
Detection: Determines the presence of materials, objects, activities, or events.
PanchromaticPanchromatic
Hyperspectral
(100’s of bands)
Hyperspectral
(100’s of bands)
MultispectralMultispectral
(10’s of bands)
MultispectralMultispectral
(10’s of bands)
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 2
Algorithms +Efficient implementations
Introduction to hyperspectral imaging: increased spectral resolution
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 3
AVIRIS (NASA/JPL) Hyperspectral Cubehttp://aviris.jpl.nasa.gov/html/aviris.freedata.html
Introduction to hyperspectral imaging: increased spectral resolution
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 4
Hyperspectral data used for demonstration:Introduction to hyperspectral imaging: increased spectral resolution
Data set provided by Robert O. Green at NASA/JPL Reference information available from U.S. Geological Survey
AVIRIS data over lower Manhattan (09/15/01) Spatial location of thermal hot spots in WTC area
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 5
Fire Temperatures
Asbestos
0
2
4
6
8
10
12
14
400 700 1000 1300 1600 1900 2200 2500
Wavelength (nm)
AVIRISEstimateResidual
WTC Hot Spot Area AHottest Spectrum
Temperature Estimate=928K6% of the area
September 11th World Trade Center
Challenges of hyperspectral data processing: computing
Debris CompositionAsbestos
AVIRIS spectra were used to measure fire temperature, asbestos contamination, and debris spread.
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 6
ContentsContents1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
Hyperspectral image analysis: contents and distribution
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009
Presence of mixed pixels in hyperspectral data
Pure pixel(water)
Mixed pixel(soil + rocks)
1000
2000
3000
4000
Ref
lect
ance
0
1000
2000
3000
4000
300 600 900 1200 1500 1800 2100 2400
Wavelength (nm)
Ref
lect
ance
Challenges of hyperspectral data processing: mixed pixels
Mixed pixel(vegetation + soil)
0
1000
2000
3000
4000
5000
300 600 900 1200 1500 1800 2100 2400
Ref
lect
ance
0300 600 900 1200 1500 1800 2100 2400
Wavelength (nm)
Wavelength (nm)
Some particularities of hyperspectral data are not to be found in other types of image data:• Mixed pixels (due to insufficient spatial resolution and mixing effects in surfaces)• Sub-pixel targets (very important and crucial in many hyperspectral applications)
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 7
Integragion of spatial and spectral information• Much effort has been given to processing hyperspectral image data in spectral terms.
• Data analysis is carried out without incorporating information about spatial context.
Pixel spatial coor-dinates randomly
shuffled
Challenges of hyperspectral data processing: integration
• There is a need to incorporate the image representation of the data in the analysis.
• Most available approaches consider spatial and spectral information separately.
• Several approaches presented in this course to achieve the desired integration.
Spectral processing Spectral processingSame output results
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 8
Why high performance computing is crucial?Biomass Burning: Sub-pixel temperatures and extent, smoke, combustion products�
Environmental Hazards: Contaminants (direct and indirect), geological substrate�
Coastal and Inland Waters: Chemical and biological standoff detection, oil spill monitoring and tracking...
Ecology: Chlorophyll, leaf water, lignin,
Biomass Burning: Sub-pixel temperatures and extent, smoke, combustion products�
Environmental Hazards: Contaminants (direct and indirect), geological substrate�
Coastal and Inland Waters: Chemical and biological standoff detection, oil spill monitoring and tracking...
Ecology: Chlorophyll, leaf water, lignin,
Challenges of hyperspectral data processing: computing
Wildland Fires in Spain/Portugal (August 2005) Imaged by MERIS sensor, European Space Agency
Ecology: Chlorophyll, leaf water, lignin, cellulose, pigments, structure, nonphotosynthetic constituents�
Commercial Applications: Mineral exploration, agriculture and forest status�
Military Applications: Detection of land mines, tracking of targets, decoys...
Others: Human infrastructure, Medical...
Ecology: Chlorophyll, leaf water, lignin, cellulose, pigments, structure, nonphotosynthetic constituents�
Commercial Applications: Mineral exploration, agriculture and forest status�
Military Applications: Detection of land mines, tracking of targets, decoys...
Others: Human infrastructure, Medical...
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 9
• Parallel computer: a collection of processing elements that cooperate to solve problems faster
• Hyperspectral imaging demands parallel computers to speed-up many applications
• Speed-up (p processors) = Performance (p processors)Performance (1 processor)
Parallel computing using commodity clustersChallenges of hyperspectral data processing: computing
Earth Simulator (5120 processors)NASA Portable MiniCluster (16 processors)
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 10
The future: onboard processing with specialized hardware
Challenges of hyperspectral data processing: computing
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 11
Challenges of hyperspectral data processing: computing
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 12
ContentsContents1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
Hyperspectral image analysis: contents and distribution
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009
Soil
Tree
Grass
Macroscopic mixture:
12 meters12
me t
ers
4 meters
4 m
eter
s
Intimate mixture:
Presence of mixed pixels in hyperspectral dataSpectral unmixing techniques: presence of mixed pixels
Macroscopic mixture:
15% soil, 25% tree, 60% grass in a 3x3 meter-pixel
Intimate mixture:
Minerals intimately mixed in a 1-meter pixel
Increasing the spatial resolution of the sensor does not necessarily solve the mixture problem
• Mixed pixls can still be obtained at very high spatial resolutions (may complicate analysis)• Intimate mixtures may take place regardless of the spatial resolution available• Models for mixed pixel characterization are required. Two strategies adopted:
ü Linear mixture model (generally unsupervised)ü Nonlinear mixture model (generally supervised, needs prior information)
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 13
Linear versus nonlinear mixing.-• Linear mixture model
ü Assumes that endmember substances are sitting side-by-side within the FOV.• Nonlinear mixture model
ü Assumes that endmember components are randomly distributed throughout the FOV.ü Multiple scattering effects.
Spectral unmixing techniques: linear versus nonlinear unmixing
Linear interaction
( ) )y,x(yx,)y,x( nMf +α= ( ) )y,x(yx,)y,x( nMf +α=
Nonlinear interaction
( )[ ] )y,x(yx, ,F)y,x( nMf +α= ( )[ ] )y,x(yx, ,F)y,x( nMf +α=
Linear mixture Nonlinear mixture
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 14
Linear spectral unmixing (LSU).-
• The goal is to find extreme pixel vectors (endmembers) that can be used to “unmix” other mixed pixels in the data using a linear mixture model.
• Each “mixed” pixel can be obtained as a combination of endmember fractional abundances. A crucial issue is how to find spectral endmembers.
1e
Spectral unmixing techniques: linear spectral unmixing
( ) )y,x(yx,)y,x( nMf +α= ( ) )y,x(yx,)y,x( nMf +α=
Band a
Ban
d b
2e
3e
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 15
Linear interaction
( ) )y,x(yx,)y,x( nMf +α= ( ) )y,x(yx,)y,x( nMf +α=
Linear mixture
Standard linear mixture-based analysis (unsupervised).-
Spectral unmixing techniques: linear unmixing methodology
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 16
Extreme pixel
Extreme pixel
Skewer 1
Skewer 2
Pixel Purity Index (PPI) algorithm.-Spectral unmixing techniques: linear unmixing methodology
Extreme pixel
Extreme pixel
Skewer 3
1) Number of ske-wers to be generated by the algorithm (k)
2) Cut-off endmem-ber threshold (t)
parameters
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 17
Spectral unmixing techniques: linear unmixing methodology
Demo: endmember extraction & unmixingDemo: endmember extraction & unmixing
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 18
Demo will be performed using ITTVIS Envi 4.5 (http://www.ittvis.com)
Nonlinear spectral unmixing.-
• The goal is to learn the properties of the mixtures in the original hyperspectral data, in order to understand nonlinearities in the process.
• We have approached this problem using supervised learning, in which highly descriptive (mixed) samples are used as input training patterns.
Spectral unmixing techniques: nonlinear spectral unmixing
( )[ ] )y,x(yx, ,F)y,x( nMf +α= ( )[ ] )y,x(yx, ,F)y,x( nMf +α=
Extreme
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 19
Nonlinear interaction
( )[ ] )y,x(yx, ,F)y,x( nMf +α= ( )[ ] )y,x(yx, ,F)y,x( nMf +α=
Nonlinear mixture Core
Border
Band 1
Ban
d 2
Spectral unmixing techniques: nonlinear spectral unmixing
Dimensionality reduction
PCA, MNF, DWT, ICA,…
Artificial neural networks for nonlinear unmixing.-• Incorporates a learning process based on known training samples:
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 20
TrainingSamples
TestSamples
Artificial Neural Network Classifier
Test classification accuracy
Randomly selected
Spectral unmixing techniques: nonlinear spectral unmixing
Architecture #1 (combined linear/nonlinear unmixing).-• Nonlinear refinement of linear abundance estimations (using FCLSU).• Limitation: this approach does not take into account the full spectral information.
p-dimensionalabundance
Estimation of number of endmembers, p
Automatic endmember
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 21
Intelligent training sample selection algorithm
Identification of the mosteffective training samples
Artificial neural network (p input and p output neurons)
Training
abundance data set
Initialcondition
Nonlinear refinement
n-dimensionalspectral data set
Automatic endmember extraction algorithm
Fully constrained linear spectral unmixing
p endmembers
Spectral unmixing techniques: nonlinear spectral unmixing
Architecture #2 (nonlinear spectral unmixing).-• MLP initially trained using only p endmembers provided by an automatic algorithm.
• Nonlinear refinement from a linear initial weight condition using training samples.
• The full spectral information is used throughout the process.
ü Choice of endmember finding algorithms for initialization
ü Choice of intelligent training sample selection algorithms
Automatic endmember
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 22
n-dimensionalspectral data set
Automatic endmember extraction algorithm
p fractional abundance
planes
p endmembers
Artificial neural network (n input and p output neurons)
Initial condition
Intelligent training sample selection algorithm
Identification of the most effective training samples
Nonlinear refinement
Mustard�s database with known proportions.-• Database of 26 pure and mixed (binary & ternary) spectra.
ü Collected using RELAB, a bidirectional spectrometer.
ü 211 spectral bands in the range 0.4 – 2.5 µm.ü Ground-truth information about fractional abundances is available for each spectra.
0,9
Olivine Enstantite Magnetite Anorthosite
Spectral unmixing techniques: nonlinear spectral unmixing
0
0,1
0,2
0,3
0,4
0,5
0,6
0,7
0,8
400 600 800 1000 1200 1400 1600 1800 2000 2200 2400
Wavelength (nm)
Reflectan
ce
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 23
True abundance proportions and selection order.-
Pure signatures
Binary mixtures
Spectral unmixing techniques: nonlinear spectral unmixing
Ternary mixtures
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 24
Abundance estimation using linear mixture model.-• Fully constrained linear spectral unmixing(FCLSU):
ü Generally accurate predictions for binary mixtures.
ü Ternary mixtures were not accurately modeled.
ü Overall RMSE scores:
Ø 11.3% (Anorthosite); 9.1% (Enstatite); 6.2% (Olivine)
Ø Unconstrained linear models always produced higher errors
Spectral unmixing techniques: nonlinear spectral unmixing
0
0,2
0,4
0,6
0,8
1
0,0 0,2 0,4 0,6 0,8 1,0
Initial estimation (Anorthosite)
Tru
e ab
unda
nce
(Ano
rtho
site
)
Anorthosite/Olivine
Anorthosite/Enstatite/Olivine
0
0,2
0,4
0,6
0,8
1
0,0 0,2 0,4 0,6 0,8 1,0
Initial estimation (Enstatite)
Tru
e ab
unda
nce
(Ens
tatit
e)
Enstatite/OlivineAnorthosite/Enstatite/Olivine
0
0,2
0,4
0,6
0,8
1
0,0 0,2 0,4 0,6 0,8 1,0
Initial estimation (Olivine)T
rue
abun
danc
e (O
livin
e)
Enstatite/OlivineAnorthosite/OlivineAnorthosite/Enstatite/Olivine
Ø Unconstrained linear models always produced higher errors
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 25
Results using architectures #1 & #2.-• Architecture #1 : Starting from linear model, samples were progressively incorporated.
ü Similar results to fully constrained linear mixture model (FCLSU)ü No significant improvements in binary and ternary mixturesü Dependent on the training samples used for modeling such mixtures
• Architecture #2 : Does not use the linear mixture model at all.ü Much better results
Spectral unmixing techniques: nonlinear spectral unmixing
ü Depends on the process of selecting training samples:
0,03
0,04
0,05
0,06
0,07
0,08
1 2 3 4 5 6 7 8 9 10
Number of training samples
RM
SE
MSAOSPMaximin
0,03
0,04
0,05
0,06
0,07
0,08
1 2 3 4 5 6 7 8 9 10
Number of training samples
RM
SE
MSAOSPMaximin
0,03
0,04
0,05
0,06
0,07
0,08
1 2 3 4 5 6 7 8 9 10
Number of training samplesR
MSE
MSAOSPMaximin
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 26
DAIS/ROSIS data collected at different altitudes.-• Obtained in of 2001 within HySens campaign of DLR at Extremadura, Spain.• Dehesa semi-arid ecosystem formed by cork-oak trees, soil and pasture.
University of Extremadura
Spectral unmixing techniques: nonlinear spectral unmixing
Guadiloba reservoir
Dehesa area
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 27
DAIS/ROSIS data collected at different altitudes.-DAIS low resolution (6 meters) DAIS high resolution (3 meters)
Spectral unmixing techniques: nonlinear spectral unmixing
ROSIS low resolution (2.4 meters) ROSIS high resolution (1.2 meters)
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 28
Reference data generation.-1. ROSIS high resolution image roughly classified into soil, cork-oak , pasture (maximum likelihood).
2. ROSIS high resolution image (1.2 m) was registered to DAIS low resolution (6 m) using an automated ground control point-based method with sub-pixel accuracy.
3. ML-generated map for ROSIS was used to provide an initial estimation of land-cover classes in the DAIS image (6x6-meter grid overlaid on 1.2x1.2-meter grid).
4. Abundance maps at the ROSIS level were thoroughly refined using field data:
a) Field spectra collected for several areas using ASD FieldSpec Pro spectro-radiometer
Spectral unmixing techniques: nonlinear spectral unmixing
b) Ground estimations of pasture abundance in selected sites of known dimensions
c) High-precision GPS work , spectral sample collection, harvest procedures, expert knowledge
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 29
Abundance estimation using linear mixture model.-• Spectral range from 504 to 864 nmwas selected for experiments (adequate for
landscape and very well covered by the two considered sensors through narrow spectral bands).
0
1000
2000
3000
504 544 584 624 664 704 744 784 824 864
Reflectance
Cork-oak treePastureSoil
Spectral unmixing techniques: nonlinear spectral unmixing
• Fully constrained linear spectral unmixing(FCLSU) better than unconstrained linear models.Wavelength (nm)
0,0
0,2
0,4
0,6
0,8
1,0
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
Cork-oak tree
RMSE=16.9%0,0
0,2
0,4
0,6
0,8
1,0
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
Pasture Soil
RMSE=9.3% RMSE=9.5%
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 30
0,0
0,2
0,4
0,6
0,8
1,0
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
0,0
0,2
0,4
0,6
0,8
1,0
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
0,0
0,2
0,4
0,6
0,8
1,0
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
0,4
0,6
0,8
1,0
Est
imat
ed a
bund
ance
s
0,4
0,6
0,8
1,0
Est
imat
ed a
bund
ance
s0,4
0,6
0,8
1,0
Est
imat
ed a
bund
ance
s
Cork-oak tree
Cork-oak tree
Pasture
Pasture
Soil
Soil
RMSE=10.3% RMSE=9.3% RMSE=9.5%
Borde:
Mezcla:
Spectral unmixing techniques: nonlinear spectral unmixing
0,0
0,2
0,4
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
0,0
0,2
0,4
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
0,0
0,2
0,4
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
0,0
0,2
0,4
0,6
0,8
1,0
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
0,0
0,2
0,4
0,6
0,8
1,0
0,0 0,2 0,4 0,6 0,8 1,0Measured abundances
Est
imat
ed a
bund
ance
s
0,0
0,2
0,4
0,6
0,8
1,0
0,0 0,2 0,4 0,6 0,8 1,0
Measured abundances
Est
imat
ed a
bund
ance
sCork-oak tree Pasture Soil
RMSE=6.1% RMSE=4.0% RMSE=6.3%
RMSE=5.9% RMSE=4.6% RMSE=4.8%
Mezcla:
Espacial:
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 31
Summary of results using nonlinear unmixing.-• How many pixel-level predictions fall within a given error bound (in percentage)
of the actual field measurements?
40%
50%
60%
70%
80%
90%
100%P
erce
ntag
e
Spectral unmixing techniques: nonlinear spectral unmixing
• Highest percentage of pixels with prediction errors below 2% achieved with mixed samples.• With the increase of error bound, spatial samples outperformed mixed samples. • Incorporation of spatial information in training sample selection can minimize the
global error, although local prediction performance may decrease in highly nonlinear classes.
10%
20%
30%
40%
2% 6% 10% 14% 18% 22% 26% 30%
Error bound
BTA MSA MEA FCLSU
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 32
Endmember extraction and unmixing practical.-Skewer 1Skewer 1
Skewer 2Skewer 2
Skewer 3Skewer 3
Extreme pixel
Extreme pixel
Extreme pixel
Extreme pixel
Extreme pixelExtreme pixel
Extreme pixelExtreme pixel
Extreme pixel
Extreme pixel
Extreme pixel
Extreme pixel
Extreme pixel
Extreme pixel
Extreme pixel
Extreme pixel
Spectral unmixing techniques: additional information & source codes
Pixel Purity Index (PPI) N-FINDR algorithm
http://www.hyperinet.eu/event1.htmlLook for “Files practice” including:
1) Matlab code for all algorithms2) Simulated hyperspectral data3) Real hyperspectral data4) Step-by-step practice session5) Exercises and questions
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 33
ContentsContents1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
Hyperspectral image analysis: contents and distribution
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009
Summary and RemarksSummary and Remarks• The special characteristics of hyperspectral images pose new processing problems, not
found in other types of remote sensing data.
• Endmember extraction and spectral unmixing allows using the increased spectral information to overcome the limitations imposed by spatial resolution (this characteristic is not found in many other types of remote sensing data)
• The integration of spatial and spectral information allows for the development of enhanced supervised/unsupervised analysis techniques.
• The special characteristics of hyperspectral images pose new processing problems, not found in other types of remote sensing data.
• Endmember extraction and spectral unmixing allows using the increased spectral information to overcome the limitations imposed by spatial resolution (this characteristic is not found in many other types of remote sensing data)
• The integration of spatial and spectral information allows for the development of enhanced supervised/unsupervised analysis techniques.
Summary and remarks
of enhanced supervised/unsupervised analysis techniques.
• Further developments are also available and expected in nonlinear spectral unmixing.
• Advances in high performance computing environments including clusters of computers and distributed grids, as well as specialized hardware modules such as field programmable gate arrays (FPGAs) or graphics processing units (GPUs) will also be crucial in many applications.
• Techniques presented in this seminar show the increasing sophistication of a field that is rapidly maturing at the intersection of many different disciplines.
of enhanced supervised/unsupervised analysis techniques.
• Further developments are also available and expected in nonlinear spectral unmixing.
• Advances in high performance computing environments including clusters of computers and distributed grids, as well as specialized hardware modules such as field programmable gate arrays (FPGAs) or graphics processing units (GPUs) will also be crucial in many applications.
• Techniques presented in this seminar show the increasing sophistication of a field that is rapidly maturing at the intersection of many different disciplines.
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009 34
HYPER-I-NET: European Research
A. Plaza, A. Mueller, R. Richter, T. Skauli, Z. Malenovsky, J. Bioucas, S. Hofer, J. Chanussot, C. Jutten, V. Carrere, I. Baarstad, P. Kaspersen, J. Nieke, K. Itten, T. Hyvarinen, P. Gamba, F. Dell’Acqua, J. A. Benediktsson, M. E. Schaepman,
J. Clevers and B. Zagajewski
HYPER-I-NET: European Research Network on Hyperspectral Imaging
Consortium
Kayser-Threde Gmbh (KT), GERMANY
Instituto Superior Técnico (IST), PORTUGAL
Institute of Systems Biology and Ecology (ISBE), CZECH REPUBLIC
Norwegian Defence Research Establishment (FFI), NORWAY
German Remote Sensing Data Center (DLR), GERMANY
University of Extremadura (UEX), SPAIN
Participating institutions:
Stefan Hofer
José Bioucas Dias
Zbynek Malenovsky
Torbjorn Skauli
Andreas Mueller
Antonio Plaza
Scientist in charge:
Bogdan ZagajewskiWarsaw University (WURSEL), POLAND
Wageningen University (WUR), THE NETHERLANDS
Norsk Elektro-Optics (NEO), NORWAY
University of Iceland (UNIS), ICELAND
University of Pavia (UNIPV), ITALY
Spectral Imaging Oy, Ltd. (SPECIM), FINLAND
Remote Sensing Laboratories, University of Zurich (UZH), SWITZERLAND
Laboratory of Planetology and Geodynamics (CNRS), FRANCE
Technical Institute of Grenoble (INPG), FRANCE
Jan Clevers
Peter Kaspersen
Jon Atli Benediktsson
Paolo Gamba
Timo Hyvarinen
Michael Schaepman
Veronique Carrere
Jocelyn Chanussot
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 2009
Main goals1. Bridge the gap between sensor design, hyperspectral data processing, and
science applications in remote sensing activities in Europe
2. Develop standardized and innovative techniques/products for hyperspectral image analysis
3. Establish standardized data processing and validation/quality 3. Establish standardized data processing and validation/quality mechanisms in all the steps of the hyperspectral processing chain
4. Integrate knowledge from different disciplines (e.g., sensor design, data processing, scientific applications)
5. Improve the cooperation and transfer of knowledge (ToK) from research centres and university groups to SMEs
6. Create a powerful multidisciplinary synergy
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 2009
Scientific areas
HYPERSPECTRAL SENSOR SPECIFICATION
Investigate the sensor requirements for various applications and develop
new sensor specifications
Coordinator: DLR
HYPERSPECTRAL PROCESSING CHAIN
Develop well-defined hyperspectral data processing chains to be used as
standardized procedures
Coordinator: University of Pavia
COORDINATION
University of
Extremadura
CALIBRATION AND VALIDATION
Calibration/validation of hyperspectral sensors and the result from steps of the processing chains
Coordinator: University of Zurich
SCIENCE APPLICATIONS
Explore relevant applications using imaging spectrometer data, and create
an application catalog
Coordinator: Wageningen University
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 2009
Network-wide training• Technical seminars for the collaboration of PhD students (arranged
to coincide with management board progress meetings)
• Secondment programme (in accordance with personal career development plans)
• Summer schools (four international joint schools)
• Summer camps (four summer camps will be held within summer schools and connected to ongoing imaging campaigns)
• Round-robin calibration experiments (during summer camps)
• Occasional short visits of studentes to other network partners
• Mid-term and final network workshops
• Training courses
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 2009
Website & e-learning
http://www.hyperinet.eu
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 2009
Website & e-learning
http://www.hyperinet.eu
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 2009
ContentsContents1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
1. Introduction to hyperspectral imaging
2. Specific challenges of hyperspectral data processing
3. Spectral unmixing techniques for hyperspectral data analysis
Hyperspectral image analysis: contents and distribution
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
4. Algorithm demonstrations
5. Summary and remarks
6. The Hyperspectral Imaging Network (Hyper-I-Net)
7. Related special issues in specialized journals
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009
Announcement: IEEE TGRS Special Issue
Call for papers IEEE Transactions on Geoscience and Remote Sensing
Special Issue on Spectral Unmixing of Remotely Sensed Data
Submission deadline: 30 September 2010 • Linear and nonlinear mixture models for analysis of remotely sensed data • Incorporation of spectral similarity measures in spectral mixture modeling • Data dimensionality issues for spectral mixture analysis • Automatic and semi-automatic endmember extraction in remotely sensed data • Supervised endmember extraction and pure class modeling • Adaptive endmember selection and multiple endmember spectral mixture analysis • Unconstrained versus constrained fractional abundance estimation in remotely sensed data • Blind source separation and its relation with spectral unmixing of remotely sensed data
Incorporation of sparsity and spatial information in spectral unmixing of remotely sensed data
Antonio J. Plaza University of Extremadura E-10071 Cáceres, SPAIN
Phone: (+34) 927 257000 (51662) Fax: (+34) 927 257202 E-mail: [email protected]
http://www.umbc.edu/rssipl/people/aplaza
Qian Du Mississippi State University
MS 39762, USA Phone: (+1) 662 325 2035 Fax: (+1) 662 325 2298
Email: [email protected] http://www.ece.msstate.edu/~du
Jose M. Bioucas Dias Technical University of Lisbon 1049-001 Lisbon, PORTUGAL Phone: (+351) 218418466 Fax: (+351) 218418472 E-mail: [email protected]
http://www.lx.it.pt/~bioucas
Xiuping Jia Australian Defence Force Academy Canberra, ACT 2600, AUSTRALIA
Phone: (+61) 2 626 88202 Fax: (+61) 2 626 88443 E-mail: [email protected]
http://www.itee.adfa.edu.au/staff/jiax/
Fred A. Kruse Arthur Brant Laboratory
University of Nevada, Reno, USA Phone: (+1) 303-499-9471 Fax: (+1) 970-668-3614 Email: [email protected]
http://www.mines.unr.edu/able/people/
• Incorporation of sparsity and spatial information in spectral unmixing of remotely sensed data • Quantitative assessment of spectral unmixing • Statistical validation of spectral mixture analysis models • Extension of spectral unmixing to multispectral scenes • Applications of spectral mixture analysis of remotely sensed data • Analysis of intimate mixtures in remotely sensed data: soil, vegetation and other application-specific studies • Spectral unmixing in planetary exploration • High performance computing implementations of spectral unmixing techniques Inquiries about the Special Issue may be directed to the Guest Editors listed below. Papers can be submitted using the manuscript central web link: http://mc.manuscriptcentral.com/tgrs and selecting Spectral Unmixing Special Issue from the ‘Manuscript type’ pull-down menu.
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009
IEEE J-STARS Special Issue Call for Papers:IEEE J-STARS Special Issue Call for Papers:
April 30, 2010
Announcement: IEEE J-STARS Special Issue
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009
Acknowledgement
AgradecimientosAgradecimientos• Proyecto European Research Netowork on Hyperspectral Imaging (MRTN-CT-2006-
035927) para formación de investigadores en análisis de imágenes hiperespectrales • Proyecto European Research Netowork on Hyperspectral Imaging (MRTN-CT-2006-
035927) para formación de investigadores en análisis de imágenes hiperespectrales
• Proyecto EODIX (AYA2008-05965-C04-02) financiado por el Ministerio de • Proyecto EODIX (AYA2008-05965-C04-02) financiado por el Ministerio de
Seminario: Avances en Espectro-Radiometría, CSIC, Madrid, 3-4 Diciembre de 2009
• Proyecto EODIX (AYA2008-05965-C04-02) financiado por el Ministerio de Educación y Ciencia y en el que participan grupos de las Universidades de Valencia, Jaume I de Castellón y Extremadura (proyecto liderado por el Prof. José Sobrino)
• Proyecto EODIX (AYA2008-05965-C04-02) financiado por el Ministerio de Educación y Ciencia y en el que participan grupos de las Universidades de Valencia, Jaume I de Castellón y Extremadura (proyecto liderado por el Prof. José Sobrino)
• Proyecto PRIPRI09A110 financiado por la Junta de Extremadura• Proyecto PRIPRI09A110 financiado por la Junta de Extremadura
• Andreas Mueller (DLR), John Mustard (Brown University) y Robert O. Green (JPL) por proporcionar los datos hiperespectrales utilizados durante el seminario
• Andreas Mueller (DLR), John Mustard (Brown University) y Robert O. Green (JPL) por proporcionar los datos hiperespectrales utilizados durante el seminario