Diapositiva 1

1
Pablo J. Martínez, Rosa M. Pérez, David Valencia,
A. J. Plaza, J. Plaza Computer Science
Department, Escuela Politécnica University of
Extremadura Avda. Universidad s/n 10071, Cáceres
(SPAIN) e-mail pablomar_at_unex.es
2
The goal of this paper is the extension of the
hyperspectral data-cube by including temporal
information, and the analysis of this new data
structure by means of spectral and temporal
endmember extraction techniques and spectral
mixture algorithms.
3
1. Change detection 2. Endmember extraction 3.
Methods 3.1 Multi temporal-spectral
signature. 3.2 Multitemporal-spectral Endmember
Extraction 4. Results and discussion 5.
Conclusions
4
(No Transcript)
5
(No Transcript)
6
Hyperspectral analysis has been a source of
innovative algorithms and techniques during the
last decade. The underlying assumption is that
the image is a matrix of spectra, moreover each
pixel vector measures the continuous spectral
response in hundreds of bands.
7

• Mixed pixels are a mixture of more than one
  distinct substances.
• Mixed pixel techniques have overcome some of the
  weaknesses of full pixel approaches by using
  mixture modelling and signal processing
  techniques.
• Spectral mixture analysis usually involves two
  steps
• Finding spectrally unique signatures of pure
  ground components (usually referred as
  endmembers)
• Expressing individual pixels in terms of linear
  combinations of endmembers.
• Classifying hyperspectral images according to
  multidimentional angles (SAM).

8
Band j
Band i
9

• Enhance the hyperspectral data-cube including
  temporal information from a sequence of images
  (a 4-D data set including the temporal
  dimension).
• Define a spectral and spatio-temporal signature.
  

10

Spectral signature of the pixel t8
t1
Spectro_temporal signature of the pixel
t8
11
Multi temporal- spectral signature
In order to detect changes in hyperspectral
signatures we must include temporal information
on this data structure. Different approach might
be used. METHOD1 Concatenating two n-
dimensional spectral signatures for different
times (t1, t2.....) we could built such data
structure.
?2n
??
?2n
?n1
??
?n
12
Bi-temporal- spectral signature
13
Bi-temporal- spectral signature
Method 2
14
MEU Algorithm

• In order to obtain a spectro_temporal
  clasification map
• Co-registration and Normalization was performed
  for t1 an t2 Images.
• Spectro-temporal (ST) image is obtained.
• Temporal Endmembers are extracted.
• Spectro-temporal image is clasified by using
  Spectral Angle Map algorithm.

15
4. Results and discussion (I)
Syntetic images
DAIS Cáceres Original image t1
Dais Cáceres Final image t2
16
Spectro-temporal endmembers
17
Temporal changes obtained by using Synthetic
images
MEU method
Ground truth
18
Confusion matrix
Ground Truth
MEU
19
Meris images
20
Ground truth image
In order to obtain a reference image for
evaluating change detection algorithms First
PCA component was obtained for t1 and t2 images.
Change detection (standardized to unit
variance) was applied to obtain a thematic change
map.
21
Ground truth image

• Bigger differences (green and red) correspond to
  clouds.
• Some geo-registration artifacts can be observed
  at the coast line and other image alignments
  (rivers, mountains.....)

Computed differences between PCA first component
Meris21-Meris30 images
22
S-t Meris endmembers
23
Comparison E.E. AMEE-SMACC
Method 1 AMEE -SAM 0.4
Method 1 SMACC SAM 0.4
24
MEU Change detection
Method 1 AMEE -SAM 0.2
Ground truth Change_detection
25
Conclusions

• MEU method provides a change image with
  information about different kinds of changes .
• Changes are easily identified using
  spectro_temporal endmembers.
• MEU provides a better background of no-change
  class and a cuanfication of changes using LSU.

Temporal unmixing