Motivation & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions A L2-Norm Regularized Pseudo-Code for Change Analysis in Satellite Image Time Series A. Radoi 1 M. Datcu 2 1 Research Center for Spatial Information (CEOSpaceTech) Dept. of Applied Electronics, University Politehnica of Bucharest 2 German Aerospace Center (DLR) LMCE 2014 First International Workshop on Learning over Multiple Contexts @ ECML A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions Motivation & Aim 1 Traditional Change Analysis Techniques 2 Pseudo-code for Change Analysis in SITS 3 Experiments 4 Conclusions 5 A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions Motivation Big Data - the actual technological developments bring large quantities of information that have to be understood and classified fast & precise Earth Observation - increasing interest in satellite image time series (SITS) A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions Motivation Big Data - the actual technological developments bring large quantities of information that have to be understood and classified fast & precise Earth Observation - increasing interest in satellite image time series (SITS) ⇒ Discover patterns of change in the temporal data A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions Motivation Big Data - the actual technological developments bring large quantities of information that have to be understood and classified fast & precise Earth Observation - increasing interest in satellite image time series (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?
Is there any difference? June 2001 October 2001 LANDSAT 7 : April 15, 1999 - still operational 16 days revisit time Our change analysis aims to: 1 reveal more than what we can learn by simply screening the images (preferably, in an unsupervised way); 2 describe the dynamic evolution of the Earth’s surface 3 keep the main properties (e.g., user-defined class) even in a time-evolving context of change.
Motivation & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions Traditional Change Analysis Techniques algebra-based techniques: image differencing and image rationing I ( t − 1) and I ( t ) two temporal images DIFF ( t ) = I ( t ) − I ( t − 1) (1) I ( t ) R ( t ) = (2) I ( t − 1) most frequently used pros: simple to implement, low complexity cons: 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 & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions Proposed Approach I ( t − 1) I ( t ) Descriptor D ( t − 1) Descriptor D ( t ) Change matrix C ( t ) λ = C λ ( D ( t − 1) , D ( t ) ) Encode change by minimizing a convex cost function: K-Means clustering Change Maps N � � J ( C ( t ) � D ( t ) − C ( t ) λ, i ⊙ D ( t − 1) 2 + λ · � d i ⊙ C ( t ) � � 2 λ, i � 2 λ ) = (3) 2 i i i =1 Images divided into N non-overlapping p × p patches ⇒ { D ( t ) i } N i =1 descriptors � � ∈ R d × N set of learned codes C ( t ) C ( t ) λ, 1 , C ( t ) λ, 2 , . . . , C ( t ) λ = λ, N A. Radoi, M. Datcu A L2-Norm Regularized Pseudo-Code for Change Analysis
Motivation & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions Datasets & Features Dataset: Landsat 7 SITS Multispectral: visible (R,G,B), near-IR (NIR), shortwave IR (SWIR 1,2) Period: 2001 – 2003 Spatial resolution: 30 meters Location: 59 × 51 km 2 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 X i ∈ R p × p → column-wise version Y i ∈ R p 2 × 1 Solve: the minimization problem n J ′ ( B , { t i } i =1 ,..., N ) = � � Y i − B · t i � 2 � � 2 + µ · � t i � 1 , (4) i =1 where B = [ B j ] j =1 ,..., d – learned dictionary t i – d - dimensional vectors that represent the projection of vector Y i onto the learned dictionary B �·� 2 and �·� 1 – L 2 - norm and L 1 - norm µ 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 ) Given: N feature points divided in: Change matrix C ( t ) = C λ ( D ( t − 1) , D ( t ) ) 4 ground-truth classes ( Water, Urban, λ Forest, Agriculture ) → { S j } 4 j =1 K clusters determined with K-Means K-Means clustering → { C k } K k =1 n k , j = | C k ∩ S j | , n k = � j n k , j , n j = � k n k , j Complete agreement or independent partitions? K Purity = 1 � j =1 ,..., |S| | C k ∩ S j | max (5) N k =1 � n k � n j � � � � � n k , j k j � 2 2 � − k , j � N 2 � 2 ARI( C , S ) = (6) � n k � n j � n k � n j � � � � � � + � � k j k j 2 2 2 2 − � N 2 � 2
Results void water forest agriculture urban (a) Image from SITS (b) Ground truth 2001 - 2002 void C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 void C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 (c) Clustering map pixel-level (d) Clustering map patch-level
Results 100 Pixels difference Pixels difference Pixels ratio Pixels ratio 0.6 95 Pixels, λ = 0.5 Pixels, λ = 0.5 Pixels, λ = 1 Pixels, λ = 1 Pixels, λ = 5 Pixels, λ = 5 90 0.5 Patches difference Patches difference Patches Ratio Patches ratio 85 Patches, λ = 0.5 Patches, λ = 0.5 0.4 Patches, λ = 1 Patches, λ = 1 Purity [%] Patches, λ = 5 ARI Patches, λ = 5 80 0.3 75 0.2 70 0.1 65 0 60 4 6 8 10 12 14 16 18 20 4 6 8 10 12 14 16 18 20 Number of clusters Number of clusters (a) Purity (b) ARI Figure : Performance measures
Motivation & Aim Traditional Change Analysis Techniques Pseudo-code for Change Analysis in SITS Experiments Conclusions Conclusions 1 Purity increases with the number of clusters ARI decreases with the number of clusters ⇒ compromise determine the optimal number of clusters 2 The proposed pseudo-encoder leads to a better separation of K-Means clusters (types of changes) 3 The method keeps the intrinsic properties as perceived by a user 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
Recommend
More recommend
