a l2-norm regularized pseudo-code for change analysis...
TRANSCRIPT
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
A L2-Norm Regularized Pseudo-Code for ChangeAnalysis in Satellite Image Time Series
A. Radoi1 M. Datcu2
1Research Center for Spatial Information (CEOSpaceTech)Dept. of Applied Electronics, University Politehnica of Bucharest
2German Aerospace Center (DLR)
LMCE 2014First International Workshop on Learning over Multiple
Contexts @ ECML
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
1 Motivation & Aim
2 Traditional Change Analysis Techniques
3 Pseudo-code for Change Analysis in SITS
4 Experiments
5 Conclusions
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
Motivation
Big Data - the actual technological developments bring largequantities of information that have to be understood andclassified fast & precise
Earth Observation - increasing interest in satellite image timeseries (SITS)
⇒ Discover patterns of change in the temporal data⇒ Data mining in change analysis
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
Motivation
Big Data - the actual technological developments bring largequantities of information that have to be understood andclassified fast & precise
Earth Observation - increasing interest in satellite image timeseries (SITS)
⇒ Discover patterns of change in the temporal data
⇒ Data mining in change analysis
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
Motivation
Big Data - the actual technological developments bring largequantities of information that have to be understood andclassified fast & precise
Earth Observation - increasing interest in satellite image timeseries (SITS)
⇒ Discover patterns of change in the temporal data⇒ Data mining in change analysis
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Is there any difference?
June 2001 October 2001LANDSAT 7 :
April 15, 1999 - still operational16 days revisit time
Our change analysis aims to:
1 reveal more than what we can learn by simply screening theimages (preferably, in an unsupervised way);
2 describe the dynamic evolution of the Earth’s surface3 keep the main properties (e.g., user-defined class) even in a
time-evolving context of change.
Is there any difference?
June 2001 October 2001LANDSAT 7 :
April 15, 1999 - still operational16 days revisit time
Our change analysis aims to:
1 reveal more than what we can learn by simply screening theimages (preferably, in an unsupervised way);
2 describe the dynamic evolution of the Earth’s surface3 keep the main properties (e.g., user-defined class) even in a
time-evolving context of change.
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
Traditional Change Analysis Techniques
algebra-based techniques: image differencing and image rationingI(t−1) and I(t) two temporal images
DIFF(t) = I(t) − I(t−1) (1)
R(t) =I(t)
I(t−1)(2)
most frequently usedpros: simple to implement, low complexitycons: not good at revealing the types of the changes
linear transformations (e.g., PCA, Tasseled Cap Transform)
classification-based methods (e.g., NN, ANN)
combinations of the above methods.
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
Proposed Approach
Encode change by minimizing aconvex cost function:
I(t−1) I(t)
Descriptor D(t−1) Descriptor D(t)
Change matrix C(t)λ = Cλ(D(t−1),D(t))
K-Means clustering
Change Maps
J(C(t)λ ) =
N∑i=1
(‖D(t)
i − C(t)λ,i �D
(t−1)i ‖22 + λ · ‖di � C
(t)λ,i‖
22
)(3)
Images divided into N non-overlapping p × p patches ⇒ {D(t)i }
Ni=1 descriptors
C(t)λ =
[C(t)λ,1,C(t)
λ,2, . . . ,C(t)λ,N
]∈ Rd×N set of learned codes
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
Datasets & Features
Dataset: Landsat 7 SITSMultispectral: visible (R,G,B), near-IR (NIR), shortwave IR (SWIR 1,2)Period: 2001 – 2003Spatial resolution: 30 meters
Location: 59 × 51 km2 around Bucharest, Romania
Features
Pixel-level: intensity of each pixel
Patch-level: sparse representation of each patch
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Learning sparse image representations
Given:
Image divided into N non-overlapping p × p patches
Each patch Xi ∈ Rp×p → column-wise version Yi ∈ Rp2×1
Solve: the minimization problem
J ′ (B, {ti}i=1,...,N) =n∑
i=1
(‖Yi − B · ti‖22 + µ · ‖ti‖1
), (4)
whereB = [Bj ]j=1,...,d – learned dictionaryti – d - dimensional vectors that represent the projection of vectorYi onto the learned dictionary B‖·‖2 and ‖·‖1 – L2 - norm and L1 - normµ models the degree of sparsity for the representation.Solution: stochastic gradient descent
Learning sparse image representations
(a) Blue filterbank (b) Green filterbank (c) Red filterbank
(d) NIR filterbank (e) SWIR1 filterbank (f) SWIR2 filterbank
Figure : Learned filterbanks from SITS
Clustering performance measures
Descriptor D(t−1) Descriptor D(t)
Change matrix C(t)λ = Cλ(D(t−1),D(t))
K-Means clustering
Given: N feature points divided in:
4 ground-truth classes (Water, Urban,Forest, Agriculture) → {Sj}4j=1
K clusters determined with K-Means→ {Ck}Kk=1
nk,j = |Ck ∩ Sj |, nk =∑
j nk,j , nj =∑
k nk,j
Complete agreement or independent partitions?
Purity =1
N
K∑k=1
maxj=1,...,|S|
|Ck ∩ Sj | (5)
ARI(C,S) =
∑k,j
(nk,j2
)−
∑k
(nk2
)∑j
(nj2
)(N2
)∑
k
(nk2
)+∑
j
(nj2
)2
−∑
k
(nk2
)∑j
(nj2
)(N2
)(6)
Results
(a) Image from SITS
void water forest agriculture urban
(b) Ground truth 2001 - 2002
void C1 C2 C3 C4 C5 C6 C7 C8 C9 C10
(c) Clustering map pixel-level
void C1 C2 C3 C4 C5 C6 C7 C8 C9 C10
(d) Clustering map patch-level
Results
4 6 8 10 12 14 16 18 2060
65
70
75
80
85
90
95
100
Number of clusters
Pu
rity
[%
]
Pixels differencePixels ratioPixels, λ = 0.5Pixels, λ = 1Pixels, λ = 5Patches differencePatches RatioPatches, λ = 0.5Patches, λ = 1Patches, λ = 5
(a) Purity
4 6 8 10 12 14 16 18 20
0
0.1
0.2
0.3
0.4
0.5
0.6
Number of clusters
AR
I
Pixels differencePixels ratioPixels, λ = 0.5Pixels, λ = 1Pixels, λ = 5Patches differencePatches ratioPatches, λ = 0.5Patches, λ = 1Patches, λ = 5
(b) ARI
Figure : Performance measures
Motivation & AimTraditional Change Analysis Techniques
Pseudo-code for Change Analysis in SITSExperimentsConclusions
Conclusions
1 Purity increases with the number of clustersARI decreases with the number of clusters⇒ compromise determine the optimal number of clusters
2 The proposed pseudo-encoder leads to a better separation ofK-Means clusters (types of changes)
3 The method keeps the intrinsic properties as perceived by auser even if the context changes over time
4 O(C ) ≈ O(DIFF ) ≈ O(R)
A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis