Transcript Noise-Robust Spatial Preprocessing Prior to Endmember Extraction

Noise-Robust Spatial
Preprocessing Prior to
Endmember Extraction from
Hyperspectral Data
Gabriel Martín, Maciel Zortea and Antonio Plaza
Hyperspectral Computing Laboratory
Department of Technology of Computers and Communications
University of Extremadura, Cáceres, Spain
Contact e-mail: [email protected] – URL: http://www.umbc.edu/rssipl/people/aplaza
Noise-Robust Spatial Preprocessing for Endmember Extraction
Talk Outline:
1. Introduction to spectral unmixing of hyperspectral data
2. Spatial preprocessing prior to endmember extraction
2.1. Spatial preprocessing (SPP)
2.2. Region-based spatial preprocessing (RBSPP)
2.3. Noise-robust spatial preprocessing (NRSPP)
3. Experimental results
3.1. Synthetic hyperspectral data
3.2. Real hyperspectral data over the Cuprite mining district, Nevada
4. Conclusions and future research lines
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
Introduction to Spectral Unmixing of Hyperspectral Data
Presence of mixed pixels in hyperspectral data
Mixed pixel
(soil + rocks)
Reflectance
4000
3000
2000
1000
0
300
600
900 1200 1500 1800 2100 2400
Pure pixel
(water)
Reflectance
Wavelength (nm)
4000
3000
2000
1000
0
300
600
900 1200 1500 1800 2100 2400
Wavelength (nm)
Mixed pixel
(vegetation + soil)
Reflectance
5000
4000
3000
2000
1000
0
300
600
900 1200 1500 1800 2100 2400
Wavelength (nm)
Some particularities of hyperspectral data not to be found in other remote sensing data:
• Mixed pixels (due to insufficient spatial resolution and mixing effects in surfaces)
• Intimate mixtures (happen at particle level; increasing spatial resolution does not address them)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
1
Introduction to Spectral Unmixing of Hyperspectral Data
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 the endmembers
e1
f ( x, y)  Mx, y   n( x, y)
Band b
Linear interaction
e3
e2
Band a
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
2
Introduction to Spectral Unmixing of Hyperspectral Data
Using spatial information in endmember extraction
• Much effort has been given to extracting endmembers in spectral terms
• Endmember extraction does not generally include information about spatial context
Pixel spatial coordinates randomly
shuffled
Endmember extraction
Same output
results
Endmember extraction
• There is a need to incorporate the spatial correlation of features in the unmixing process
• We develop a new strategy to include spatial information in endmember extraction
• The method works as a pre-processing module (easy to combine with available methods)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
3
Noise-Robust Spatial Preprocessing for Endmember Extraction
Talk Outline:
1. Introduction to spectral unmixing of hyperspectral data
2. Spatial preprocessing prior to endmember extraction
2.1. Spatial preprocessing (SPP)
2.2. Region-based spatial preprocessing (RBSPP)
2.3. Noise-robust spatial preprocessing (NRSPP)
3. Experimental results
3.1. Synthetic hyperspectral data
3.2. Real hyperspectral data over the Cuprite mining district, Nevada
4. Conclusions and future research lines
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
Spatial Preprocessing Prior to Endmember Extraction
Spatial Pre-Processing (SPP)
Developed by Zortea and Plaza (IEEE Trans. Geosci. Remote Sens., 2009)
1.
Move a spatial kernel around each hyperspectral pixel vector and calculate a spatial
correction factor for each pixel
2.
Assign a weight to the spectral signature of each pixel depending on the spectral
similarity between each pixel and its spatial neighbors, so that anomalous pixels are
displaced to the centroid, while spatially homogeneous pixels are not displaced
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
4
Spatial Preprocessing Prior to Endmember Extraction
Spatial Pre-Processing (SPP)
Developed by Zortea and Plaza (IEEE Trans. Geosci. Remote Sens., 2009)
1.
Move a spatial kernel around each hyperspectral pixel vector and calculate a spatial
correction factor for each pixel
2.
Assign a weight to the spectral signature of each pixel depending on the spectral
similarity between each pixel and its spatial neighbors, so that anomalous pixels are
displaced from the centroid, while spatially homogeneous pixels are not displaced
3.
Apply spectral-based endmember extraction (using, e.g., OSP, VCA or N-FINDR) after
the preprocessing, obtaining a final set of endmembers from the original image
Band Y
e1
e3
e2
Band X
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
4
Spatial Preprocessing Prior to Endmember Extraction
Region-Based Spatial Pre-Processing (RBSPP)
Developed by Martín and Plaza (IEEE Geosci. Remote Sens. Lett., 2011)
Estimation of the number
of endmembers p
Hyperspectral
image with n
spectral bands
Several possibilities: Chang’s VD;
Bioucas’ HySime; Luo and
Chanussot’s eigenvalue approach
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
5
Spatial Preprocessing Prior to Endmember Extraction
Region-Based Spatial Pre-Processing (RBSPP)
Developed by Martín and Plaza (IEEE Geosci. Remote Sens. Lett., 2011)
Estimation of the number
of endmembers p
Hyperspectral
image with n
spectral bands
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
Unsupervised
clustering
ISODATA is used to partition the
original image into c clusters, where
cmin=p and cmax=2p
5
Spatial Preprocessing Prior to Endmember Extraction
Region-Based Spatial Pre-Processing (RBSPP)
Developed by Martín and Plaza (IEEE Geosci. Remote Sens. Lett., 2011)
Estimation of the number
of endmembers p
Unsupervised
clustering
Hyperspectral
image with n
spectral bands
Morphological erosion and
redundant region thinning
Intended to remove mixed pixels at
the region borders; multidimensional
morphological operators are used to
accomplish this task
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
5
Spatial Preprocessing Prior to Endmember Extraction
Region-Based Spatial Pre-Processing (RBSPP)
Developed by Martín and Plaza (IEEE Geosci. Remote Sens. Lett., 2011)
Estimation of the number
of endmembers p
Unsupervised
clustering
Region selection using
orthogonal projections
Morphological erosion and
redundant region thinning
Hyperspectral
image with n
spectral bands
An orthogonal subspace projection
approach is then applied to the
mean spectra of the regions to
retain a final set of p regions
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
5
Spatial Preprocessing Prior to Endmember Extraction
Region-Based Spatial Pre-Processing (RBSPP)
Developed by Martín and Plaza (IEEE Geosci. Remote Sens. Lett., 2011)
Estimation of the number
of endmembers p
Unsupervised
clustering
Region selection using
orthogonal projections
Morphological erosion and
redundant region thinning
Hyperspectral
image with n
spectral bands
Preprocessing module
Automatic endmember
extraction and unmixing
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
p fully constrained abundance maps (one
per endmember)
5
Spatial Preprocessing Prior to Endmember Extraction
Noise-robust spatial preprocessing (NRSPP)
•
The method first derives a spatial homogeneity index which is relatively insensitive to
the noise present in the original hyperspectral data; then, it fuses this index with a
spectral-based classification, obtaining a set of pure regions which are used to guide the
endmember searching process
•
Step 1: Apply multidimensional Gaussian filtering using different scales, which
results in different filtered versions of the original hyperspectral image
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
6
Spatial Preprocessing Prior to Endmember Extraction
Noise-robust spatial preprocessing (NRSPP)
•
Step 2: Calculate the root mean square error (RMSE) between the original
image and each of the filtered images and derive a spatial homogeneity
index as the average of the obtained difference values; such spatial
homogeneity calculation is robust in the presence of noise
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
7
Spatial Preprocessing Prior to Endmember Extraction
Noise-robust spatial preprocessing (NRSPP)
•
Step 3: Perform a spectral-based unsupervised classification of the original
image; here, we use the ISODATA algorithm, where the number of
components retained was set to p, the number of endmembers
•
Step 4: For each cluster in the classification map, a percentage (alpha) of
spatially homogeneous pixels are selected; then, we apply the OSP
algorithm over the averaged signatures in each resulting region to select the
most highly pure regions (removing those which contain mixed pixels)
•
Endmember extraction is finally applied to the pixels retained after the
NRSPP, which acts as a pre-processing module (as the SPP and RBSPP)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
8
Noise-Robust Spatial Preprocessing for Endmember Extraction
Talk Outline:
1. Introduction to spectral unmixing of hyperspectral data
2. Spatial preprocessing prior to endmember extraction
2.1. Spatial preprocessing (SPP)
2.2. Region-based spatial preprocessing (RBSPP)
2.3. Noise-robust spatial preprocessing (NRSPP)
3. Experimental results
3.1. Synthetic hyperspectral data
3.2. Real hyperspectral data over the Cuprite mining district, Nevada
4. Conclusions and future research lines
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
Experimental Results with Synthetic and Real Hyperspectral Data
Synthetic Image Generation
•
The scenes have been generated using fractals to generate random spatial patterns
•
Each fractal image is divided into a set of classes or clusters
•
Mixed pixels are generated inside each cluster using library signatures
•
Spectral signatures obtained from a library of mineral spectral signatures available online
from U.S. Geological Survey (USGS) – http://speclab.cr.usgs.gov
•
Random noise in different signal-to-noise ratios (SNRs) is added to the scenes
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
9
Experimental Results with Synthetic and Real Hyperspectral Data
Synthetic Image Generation
•
Database available online: http://www.umbc.edu/rssipl/people/aplaza/fractals.zip
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
10
Experimental Results with Synthetic and Real Hyperspectral Data
Experiments with Synthetic Images
•
Average spectral angle (degrees) between ground-truth USGS spectra and the
endmembers extracted across five synthetic scenes with different SNRs (alpha=70)
•
RMSE after reconstructing the five synthetic scenes (with different SNRs) using
the endmembers extracted by OSP (alpha=70)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2011), Vancouver, Canada, July 24 – 29, 2011
11
Experimental Results with Synthetic and Real Hyperspectral Data
AVIRIS Data Over Cuprite, Nevada
IEEE International Geoscience and Remote Sensing Symposium (IGARSS’2011), Vancouver, Canada, July 24 – 29, 2011
12
Experimental Results with Synthetic and Real Hyperspectral Data
Experiments with the AVIRIS Cuprite hyperspectral image
RMSE=0.165
OSP (81 seconds)
RMSE=0.129
SPP+OSP (49+81 seconds)
RMSE=0.101
RMSE=0.265
AMEE (96 seconds)
SSEE (320 seconds)
RMSE=0.085
RMSE=0.067
RBSPP+OSP (78+14 seconds)
NRSPP+OSP (71+12 seconds)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS’2011), Vancouver, Canada, July 24 – 29, 2011
Times
measured in
Intel Core i7
920 CPU at
2.67 GHz with
4 GB OF RAM
(p = 22)
13
Conclusions and Hints at Plausible Future Research
Conclusions and Future Lines.•
We have developed a new spatial pre-processing method which can be used prior
to endmember extraction and spectral unmixing of hyperspectral images
•
The proposed method shows some advantages over other existing approaches, in
particular, when the noise level in the hyperspectral data is relatively high
•
The results obtained with synthetic scenes anticipate that the incorporation of
spatial information may be beneficial in order to allow a better modelling of
spatial patterns and robustness in the presence of noise
•
The results obtained with real scenes indicate that the incorporation of spatial
information directs the endmember searching process to spatially
homogeneous regions in the original hyperspectral scene
•
Future work will be directed towards comparisons with multiple endmember
spectral mixture analysis techniques (comparable in terms of abundance
estimation accuracy but more complex in computational terms)
IEEE International Geoscience and Remote Sensing Symposium (IGARSS’09), Cape Town, South Africa, July 12 – 17, 2009
14
IEEE J-STARS Special Issue on Hyperspectral Image and Signal Processing
IEEE International Geoscience and Remote Sensing Symposium (IGARSS’09), Cape Town, South Africa, July 12 – 17, 2009
15
Noise-Robust Spatial
Preprocessing Prior to
Endmember Extraction from
Hyperspectral Data
Gabriel Martín, Maciel Zortea and Antonio Plaza
Hyperspectral Computing Laboratory
Department of Technology of Computers and Communications
University of Extremadura, Cáceres, Spain
Contact e-mail: [email protected] – URL: http://www.umbc.edu/rssipl/people/aplaza