P' Blonda, C' Tarantino

1
Automatic thematic change detection

• P. Blonda, C. Tarantino
• Istituto di Studi sui Sistemi Intelligenti per
  lAutomazione
• ISSIA-CNR
• Via Amendola, 166/5 -70126 Bari
• blonda_at_iesi.ba.cnr.it

2
Outline

• Introduction
• Unsupervised change detection
• Supervised change detection
• Hybrid techniques
• Future work
• ground truth availability
• input features (spectral-spatial-temporal) and
  images
• performance evaluation
• limits and advantages
• references

3
Introduction

• EO data offer the opportunity to automatically
  detect thematic changes useful in long term
  studies (ecosystem-monitoring) as well as in
  quasi real-time disaster monitoring. Focus is
  on
• the place where changes occurred
• the specific class transitions involved
• when they occurred time series analysis.
• First requirement a model for the natural or
  man induced process (or the event ) to be
  monitored trough change detection
• Second requirement evaluation of the global
  change detection procedure

4
Data and methodology selection

• From the model
• specific classes of interest and
  class-transitions
• frequency of change map production
• one per year, one per month, every-day, in the
  case of disaster, one after disaster and some
  days (months) after
• satellite images (spectral-spatial resolution vs.
  scale of the process)
• ground truth data and other ancillary data (DEM,
  pluviometric , prior knowledge on the study area)
• 
  
• methodoly selection
• after limits and advantages evaluation,
• work hypotesis explicitation
• false change reduction

5
Test sites and change detetion techniques versus
available ground truth

• Caramanico test site
• TM images, 1986-2000
• CORINE land use/cover maps 2000
• ortophoto 1997.
• Daunia test site
• TM images, 1986-2001
• CORINE 2000 and ortophoto 1997 and 2003.
• Greek test site
• TM images, 1995-2001
• some pixels of known position (GPS data)
• and ground truth acquired in 2005.

EVG1-CT-2001-00055 LEWIS
Supervised a) data driven by unchanged ROIs
knowledge driven
unsupervised
hybrid compound
6
Thematic Change Detection

• Techniques
• Post Classification Comparison (PCC)
  independent classification
• Supervised data-driven classifiers
• Prior knowledge-driven classifiers (Baraldi 2006)
• Cascade Compound Classification (Bruzzone
  97,99) temporal correlation
• Hybrid techniques
• hybrid (unsupervised/supervised) compound
  (Castellano, 2006)
• Two-stage classifier (knowledge-driven/data
  driven ) (Baraldi, 2006)
• Input Features
• spectral and spatial (context as GMRF,
  Co-occurence matrix)
• geostatistics ( Fiorentino, 2006) ( Boucher,
  2006).

7
Quantitative Performance Evaluation

• Given two images X1 and X 2 at time t1 and t2,
  respectively
• In PCC from independent classification, given
  OA1, OA2
• The percentage Correct Classification
• The Jaccard Coefficient
• The Yule Coefficient
• Given a time series, change detection accuracy
  is calculated as the percentage of pixels having
  the vector of label classes all correct

on test data
8
A.1 PCC by independent data-driven classification
CLASSIFIER SELECTION Mutually exclusive,
i.e. crisp thematic maps are generated.
Non-mutually exclusive, i.e. soft thematic maps
are generated. Totally exhaustive, i.e. there is
no class unknown. Non-totally exaustive, i.e.
featuring class unknown.
B. ROIs selection and feature extraction
A. Pre-processing
Selected
Enhanced
A
B
Raw Images at T1 and T2
features
images
D. Post Classification Comparisoc (PCC)
C. Classification
Map of
Classified
D
C
change
map
PCC does not take into account the temporal
dependence beteen two acquisitions of the same
area
E. Accuracy assessment
9
Change detection by independent classification

• 1) Generation of thematic maps from each date.
• 1.a) Posterior probability values.
• 1.b) Confusion matrices on test data for
  evaluation.
• 2) Classification improvement
• by thresholding the posterior probabilities
  values
• by modifying the classifier (topology,
  parameters, learning rule, input features)
•  3) Comparison of year-specific thematic map
  pairs.
• 3.a) Change images and change matrices
  from pairs of thematic maps.
• 3.c) Joint probability and
  class-conditional probability estimates.
• 3.d) Temporal trends of class-specific
  occurrences.

GMRF Melgani 2003 Serpico, 2006 SVMcontext Bruz
zone , 2006
10
Change detection by independent classification
From change matrix, by using Bayesian rule, the
transition probabilities are evaluated, i.e.
class-conditional probabilities as
P2-gt1(i / k)
P2-gt1(i , k) / P1(k) Nik/ N1k where

P2-gt1(i , k) N i , k/ N
joint probability P2(i) ?k  N i, k / N
N2i / N prior probability of class i at t t
2 P1(k) ?i N i, k / N N1k/ N prior
probability of class k at t t 1 Let T 2lt-1 be
the transition probability matrix with components
P (i k) for i,k 1,, NC. Let P t be the
vector of prior probability at time t with
components Pt(m), m 1,..,NC. It contains
information on the number of pixels assigned to
each class at time t When vector P 1 is known,
then P t at t 2 can be computed as
 P 2 T2-gt1 P 1 P 3
T 3-gt2 P 2 T 3lt-2 T2lt-1 P 1
11
Example spectralspatial trough geostatistics
Indicator kriging
Given a set of labeled pixels in the image,
denoted as Bgeo each
pixel in the image must be classified into one of
L classes by using the information Bgeo. The
indicator kriging is an optimal linear estimator
that can provide the probability
The variorgam for each class,
At each location u, the set of L probability must
satisfy
It derives from the minimization of the error
between the espected and true values
It derives from the non-distorsion condition
12
Spatial information by geostatistics
C. Fiorentino, C. Tarantino, A. Castrignano, G.
Pasquariello, Use of geostatistical analysis to
improve classification, Proc. SPATIAL Conference,
Foggia, Italy, 2006.
13
A.2) Supervised Knowledge Based system

• The problem Surface areas to be analyzed are
  large and heterogeneous, and it is often too
  expensive to collect ROIs to estimate the
  statistics of every target class.
• The proposed system maps each pixel data vector
  into a finite set of discrete spectral categories
  (kernel) .
• To detect kernel spectral categories (fuzzy)
  decision rules combine several sources of
  spectral evidence taken from literature.
• Baraldi, V. Puzzolo, P. Blonda, L. Bruzzone, C.
  Tarantino, Automatic spectral rule based
  preliminary mapping of calibrated Landsat TM and
  ETM images IEEE Trans.on Geosc. and Remote
  Sensing, Vol.44, no 9, Sept 2006.

14
Caramanico supervised change detection
Ortophoto 1997
Co-registered images
15
Regione Abruzzo test site, Caramanico area
Lansat TM May 1997, RGB5-4-1
Kernel spectral category map from KB system
16
Resuts of Landslide Early Warning Intergrated
System
Class distribution in time on the Landslide area
17
Resuts on two sub-windows in the valley
18
Compound classification
B. ROI selection and feature extraction
A. Preprocessing
Selected
Raw
Enhanced
A
B
feature
image
image
image
D. Compound Classification
C. Independent Classification
Classified maps C1 and C2
D
C
E. Map comparison
Number of classes and ROIs
Temporal Information
E
Time dependency between the a priori
probabilities of classes
Bruzzone, 1997, 1999, 2002.
19
Bayesian rule
Compound Decision rule
or
20
Exampe Spectral -Temporal and Spatial trough
geostatistics

• Spatial information is considered by introducing
  in the prior probability the result indicator
  kriking.
• Data used a time series of 6 TM images between
  1988 and 1996
• Ground truth 1971 locations known at all times
  and corresponding to 7 classes water, forests,
  agriculture, urban, fish pond, shrub, transition

Average OA improved marginally from 78 to
82 Time series accuracy improved 33 to 61
most of the incorrect classifications are
within a reduced number of images
(filtering). Reduction of unwanted speckles in
the classified image.
A. Boucher, K. C. Seto, A. G. J., A Novel Method
for Mapping Land Cover Changes Incorporating
Time and Space With Geostatistics, IEEE
Transaction on Geoscience and Remote Sensing Vol.
44, No. 11, pp. 3427-3435, Nov. 2006.
21
The methodology
temporal information
spectral
spatial
22
t model to combine the probabilitythe
probabilities are expressed as distances related
to the likelihood of event L(u,t)k as follows
The posterior probability is then retrieved by
inverting from the updated distance
with tiso and tT parameters measuring redundacy
between information sources
23
Unsupervised change detection
A. Preprocessing co-registration and geometric
correction
A
Enhanced
Input raw images T1 and T2
Images T1 andT2
B. Change Vector Anaysis (CVA)
C. Thresholding
Difference Image
C
Change map and percentage of changed pixels

• B
• direct spectral values comparison
• clustering weights comparison

• They strongly depends on data pre-processing
  for radiometric and atmospheric effects.
• Advantage no ground truth is required for data
  processing but only for validation

24
Unsupervised change detection

• B. Extraction of a difference image
• Univariate Image Differencing (one band is
  considered)
• Change Vector Analysis (CVA) (difference of the
  feature vectors)
• the statistical analysis of the magnitudes
  allows to detect changes
• the directions is used to distinguish among
  different transitions
• Vegetation index differencing
• Other linear or non-linear combination bands
  differencing
• Pricipal Component Analysis
• Image clustering (NNs) and weights comparison

25
Unsupervised change detection

• Limits are mainly realated to differences in
• atmospheric conditions at two times
• different phenological states
• sensor calibration
• moisture conditions
• Advantages
• no ground truth is required
• Best global threshold selection in the
  Difference image histogram
• Automatic (Bruzzone 2000) based on mixture
  modelling
• Combination of decision of an ensemble of
  different thresholding algorithms trough MRFs
  (Melgani, 2006)

26
Hybrid compound classification
L. Castellana et al., Pattern
Recognition Letters, 2006, in press.
A. Preprocessing
B. ROI selection and feature extraction
Selected
Raw
Enhanced
A
B
feature
images
images
image
D. Compound Classification
C. Independent Classification
Classified maps C1 and C2
D
C
E. Map comparison
Number of classes and ROIs
E
U. Unsupervised change extraction
U
Prior information is extracted by unsupervised
change extraction
27
Test site Caramanico in Abruzzo (I)
Landsat TM Two images per year from 1986- 2000
28
Caramanico test site
Window a
CVA difference image between pre-processed
images, T1 and T2, of the same month
Change image after automatic thresholding
29
a
30
Unsupervised technique results
31
TD_PCC Transition probability values
? ( no-change percentage) iff
iff
since the constraint
has to be satisfied
32
TD_PCC Transition probability values
33
b
b
34
Greek test site Prior knowledge-based
classification

• 2 landslide test sites Pititsa and Panagopoula
• 1. 1975 Nov-Dec. 1998 March 2002
• 2. 1971-June 1995 (2)

• SPOT4 land-use map dated 1995
• 25 labeled pixels, located by GPS in 2003.
• DEM

35
October 1995 72 cathegories by knowledge based
classifier
Test site
36
october 1992 RGB 5-4-1 october 1995 RGB
5-4-1 october 1998 RGB 5-4-1
Greek test site unsupervised rule-based
classification
may 2001 RGB 4-3-2 may 2001
RGB 4-5-6 october 2001 RGB 5-4-1
37
october 1992 october 1995
october 1998

• may 2001 october
  2001

38
(No Transcript)
39
Conclusions and future work on CD

• Unsupervised
• They strongly depend on data pre-processing for
  radiometric and atmospheric effects but no ground
  truth is required
• automatic threshold selection
• neural networks (Pajares, 2006)

• Supervised Data Driven classification based CD
• Input features integration (spectral-spatial-tempo
  ral casacade/compound )
• Ground truth by (Landgreebe, Bruzzone 2002)

• Prior-Knowledge based classification CD
• Automatic extraction of kernel spectral maps
  useful for photo interpreters
• in ROIs selection (to be used in hybrid CD
  systems).
• Multitemporal rule definition for Corine level
  4/5 classification
• Rule for hyper-spectral data classification

• Hybrid change detection system
• Data fusion (unmixing from hyperspectral data,
  medium resolution data)
• Decision fusion from different data processing
  modules

40
Requirements for KIM

• Data
• banchmark image data sets,
• ancillary data for test area
• prior knowledge on the area (web, news papers)
• Algorithms (assessed or to be assessed in terms
  of generalization trough application to different
  data sets)
• Previous results
• Common evaluation tools
• Data and algorithms policy

41
References

• F. Melgani, and Y. Bazi, Markovian Fusion
  Approach to Robust Unsupervised Change Detection
  in Remotely Sensed Imagery, IEEE Geosc. Remote
  Sensing Letters VOL. 3, NO. 4, pp. 457-461, OCT.
  2006.
• C. Tarantino, P. Blonda, G. Pasquariello, Remote
  Sensed Data for Automatic Detection of Land Use
  Changes due to Human Activity in Support to
  Landslide Studies, Natural Hazards, November (on
  line) 2006.
• Qiong Jackson, and David A. Landgrebe, Adaptive
  Bayesian Contextual Classification Based on
  Markov Random Fields, IEEE Transaction on
  Geoscience and Remote Sensing, Vol. 40, No. 11,
  pp. 2454-2463, Nov. 2002.
• S. B. Serpico, G. Moser, Weight Parameter
  Optimization by the HoKashyap Algorithm in MRF
  Models for Supervised Image Classification, IEEE
  Transaction on Geoscience and Remote Sensing,
  Vol. 44, no. 12, pp. 3695-3705, Dec. 2006.
• L- Bruzzone, M. Chi, M. Marconcini, A Novel
  Transductive SVM for Semisupervised
  Classification of Remote-Sensing Images, IEEE
  Transaction on Geoscience and Remote Sensing,
  Vol. 44, no. 12, pp. 3363-3373, Dec. 2006.
• Gonzalo Pajares, A Hopfield Neural Network for
  Image Change Detection, , IEEE Transaction on
  Geoscience and Remote Sensing, Vol. 17, no. 5, ,
  pp. 1250-1264, Sept. 2006.
• Asbjørn Berge, and Anne H. Schistad Solberg,
  Structured Gaussian Components for Hyperspectral
  Image Classification, IEEE Transaction on
  Geoscience and Remote Sensing Vol. 44, No. 11,
  pp. 3386-3396, Nov. 2006.
• Suju Rajan, Joydeep Ghosh, and Melba M. Crawford,
  Exploiting Class Hierarchies for Knowledge
  Transfer in Hyperspectral Data, IEEE Transaction
  on Geoscience and Remote Sensing, Vol. 44, No.
  11, pp. 3408- 3417, Nov. 2006.