IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1 Interactive …sahbi/grsl2017.pdf · 2017-09-21 · IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1 Interactive Satellite Image Change Detection

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1

Interactive Satellite Image Change Detection

with Context-Aware Canonical Correlation

AnalysisHichem Sahbi, Member, IEEE

Abstract

Automatic change detection is one of the remote sensing applications that has received an increasing attention

during the last years. However, fully automatic solutions reach their limitation; on the one hand, it is difficult to

design general decision criteria able to select area of changes for images under various acquisition conditions, and

on the other hand, the relevance of changes may differ from one user to another.

In this paper, we introduce an alternative change detection method based on relevance feedback. The proposed

algorithm is iterative and based on a query and answer (Q&A) model that (i) asks the user questions about the

relevance of his targeted changes, and (ii) according to these answers updates change detection results. Our method

is also based on a new formulation of canonical correlation analysis (CCA), referred to as context-aware CCA, that

learns transformations which map data from different input spaces (related to multi-temporal satellite images) into

a common latent space which is sensitive to relevant changes while being resilient to irrelevant ones. These CCA

transformations correspond to the optimum of a particular constrained maximization problem that mixes an alignment

term with a context-based regularization criterion. The particularity of this novel CCA approach, resides in its ability

to exploit spatial geometric context resulting into better performances compared to other CCA approaches, as shown

in experiments.

Index Terms

Satellite image change detection, canonical correlation analysis, resilience, context, relevance feedback

I. INTRODUCTION

Change detection is the process of identifying occurrences of targeted changes into a scene at a given instant t1

w.r.t the same scene, acquired at t0 < t1. In remote sensing satellite imagery, acquisitions may be of different natures

and applications are numerous ranging from studying environmental variations (melting glacier, deforestation,

etc.) [1], [2], [3], to assessing damaged areas after catastrophe (flooding, earth-quakes, fires, tornados, etc.) [4], [5],

[6], to video surveillance and mapping services. Early change detection algorithms were initially based on simple

comparisons of multi-temporal signals, via image differences and thresholding, using vegetation indices, principal

component and change vector analyses, operating either at the pixel or the object levels (see [7], [8], [9], [10], [11],

Manuscript received June 23, 2016; revised October 5, 2016; accepted December 26, 2016.


[12], [13] and references therein).

Depending on use-cases, one may identify relevant changes (appearance or disappearance of entities/objects into

scenes either continuously or discontinuously in space/time) and many irrelevant changes (due to radiometric and

atmospheric variations, shadow, registration errors, insignificant local motions of objects like waving trees, etc.); for

instance, in damage assessment after natural catastrophes, relevant changes – for some users/operators – might be

areas of building destructions and flooded zones, while for other users relevant changes might be obstacles in roads.

Existing change detection methods (see for instance [14], [15], [16], [17], [18]) rely on a preliminary preprocessing

step that attenuates the effects of irrelevant changes, by finding parameters of sensors for registration as well as

correcting radiometric effects, occlusions and shadows [19], [20]. Other methods [21], [22], [23], [24], [25], [26],

[27] either ignore irrelevant changes or consider them as a part of appearance model design and are able to detect

relevant changes while being resilient to irrelevant ones.

Among automatic change detection solutions, correlation-based methods (such as CCA) are particularly successful

(see for instance [28], [29], [30]), but their success is highly dependent on the following hypothesis: information

(pixels, patches, etc.) from the same physical areas between multi-temporal images are highly correlated when

irrelevant changes occur, and uncorrelated with relevant changes. However, this hypothesis does not stand all the

time in practice; on the one hand, change detection methods operating at the pixel-level (and even at the patch-

level), are powerless to capture sufficiently discriminating information in order to precisely decide whether a pixel

or a patch belongs to a “change” or “no-change” class. Besides, pixel and patch-based methods require accurate

alignments through multi-temporal images which may be difficult to obtain as pixel and patch intensities are very

sensitive to irrelevant changes including noise and radiometric variations. On the other hand, the frontier between

relevant and irrelevant changes is sometimes narrow and may differ from one user to another, and hence requires

the intervention of users to designate a few examples of positive and negative changes according to their intentions,

in order to obtain accurate change decision criteria.

In this paper, we tackle these issues and we propose a change detection algorithm, for satellite images, based

on relevance feedback (RF), a.k.a interactive image search, and a novel CCA method. The latter, referred to as

context-aware CCA, models the statistical correlation between satellite image data (taken at different instants) and

robustly map image regions (belonging to the same physical locations) from input images to latent spaces while

being sensitive to relevant changes and resilient to irrelevant ones. Our contribution considers the spatial context as

a part of CCA design and this produces significantly better performances, especially when the spatial coherence of

changes is relevant1. The formulation is based on the optimization of an objective function mixing two terms; the

first one partly similar to standard CCA seeks to maximize the correlation between aligned “no-changes” in satellite

image data (taken at different instants and under different acquisition conditions) while at the same time it reduces

the correlation between “changes”. The second term is a regularization criterion that diffuses the correlations,

through context, between neighboring data, resulting into an extra gain as shown through experiments in change

1when changes are due to natural hazards such as tornados, even though their first impact points are difficult to predict precisely, tornados

usually follow smooth paths so changes usually spread smoothly in areas impacted by these phenomena.


detection.

II. OUR RF-BASED CHANGE DETECTION AT A GLANCE

Define I0 = {u1, . . . ,un}, I1 = {v1, . . . ,vn} as two satellite image patches (referred to as reference and test

images resp.) captured at instants t0 and t1 (t0 < t1). Consider I = {x1, . . . ,xn} as a set of registered patch

pairs, with xi = (ui,vi) ∈ R2d and Y = {y1, . . . ,yn} the underlying unknown labels. Our goal is to design

an RF-based change detection algorithm which predicts the unknown labels {yi}i with yi = +1 if the test patch

vi ∈ I1 corresponds to a “change” w.r.t its reference patch ui ∈ I0; and yi = −1 otherwise. Let Dt ⊂ I be a

display (subset of patch pairs with |Dt| = 16� |I| in practice) shown to an oracle (user) at iteration t and let Ytbe the unknown labels of Dt. Starting from a random D0, including representative samples in I, we update our

change detection criteria by running the following steps for t = 0, . . . , T − 1 (In practice, T = 10); see Fig. 1

Step 1 (Oracle model). Label display Dt with a known-only-by-the-user oracle function (denoted C(.)) and assign

C(Dt) to Yt. As our change detection ground-truth is objective, we assume deterministic oracle functions only.

Step 2 (Learning model). Train a classifier ft(.) on data labeled, so far, ∪tk=0(Dk,Yk) to predict unknown labels

in I − ∪tk=0Dk. In practice, we use LIBSVM with the triangular kernel [31], in order to build ft(.); this choice

was motivated by the good performance of RF when using the triangular kernel compared to many other kernels

(see [31], [32]). Note that SVMs are trained on top of novel context-aware CCA features, and this is the main

contribution of this work.

Step 3 (Display model). Select the next display Dt+1 ⊂ I − ∪tk=0Dk to label by the oracle using two strategies,

closely related to active learning [33]: exploration and exploitation. The former selects data to discover new modes of

our change detection criteria {ft}t while the latter locally refines these criteria. Our strategy seeks a balance between

exploration and exploitation: at t = 0, we apply exploration, then at each iteration t ≥ 1, we select the subsequent

display Dt+1 depending on how good was the previous one; i.e., we either keep the previous action (exploration

or exploitation) or we switch from one to another, depending on a score St =∑

x∈Dt1{sign[ft−1(x)] 6=C(x)} which

measures how informative is display Dt obtained using the previous action2.

III. ENHANCING THE LEARNING MODEL WITH CCA

Standard CCA (e.g., [34]) finds two transformations that map aligned data {(ui,vi)}i in I0 × I1 into a latent

space while maximizing their correlation. Let Pu, Pv denote these transformations which respectively correspond to

reference and test images. CCA finds these transformations by maximizing P′vCvuPu, subject to P′uCuuPu = 1,

P′vCvvPv = 1; here ′ stands for the transpose operator, Cvu (resp. Cuu, Cvv) correspond to the interclass

(resp. intraclass) covariance matrices of data in I0, I1, and equality constraints control the effect of scaling on the

solution. One can show that problem above is equivalent to solving the generalized eigenproblem CuvC−1vv CvuPu =

γ2CuuPu with Pv =1γ C−1vv CvuPu. In practice, learning these two transformations requires labeled “no-changes”

in I. A variant in [30], referred to as discriminant CCA, combines both “no-changes” and “changes” discriminatively

2a good action should produce a display to correct as many change detection results as possible thereby better refining the subsequent decision

criterion.


Satellite Image

Data

Oracle

Learning

CCA+SVM

New

Display (Q

uestions)

Initial Display

Display

ModelCCA features + Classifier

Fee

dbac

k (A

nsw

ers/

Lab

els)

Fig. 1. Flowchart of our iterative and interactive change detection algorithm.

Reference patches Test patches

latent space (CA CCA)

Pu

φv(vi)

φu(ui)

Pv

viui

Reference patches Test patches

latent space (CA CCA)

Pu Pv

ui vi

φu(ui)

φv(vi)

Fig. 2. This figure shows the diffusion of correlations between neighboring patch pairs when using context-aware CCA; the thickness of the

dashed blue lines corresponds to the amount of patch correlations between reference and test images. (Left) Though patch pairs (ui,vi) are

initially weakly correlated, they become highly correlated in the latent space as their contextual patches (i.e., immediate neighbors) are highly

correlated. (Right) In contrast, though patch pairs (ui,vi) are initially highly correlated, they become weakly correlated in the latent space as

their contextual patches are weakly correlated.

and provides better results compared to standard CCA (see also Section IV); precisely, discriminant CCA seeks

to maximize P′vC−vuPu − λP′vC

+vuPu with λ ∈ R+ and C−vu (resp.C+

vu) being the covariance matrix of negative

(resp. positive) data in {(ui,vi)}i. However, using only labeled patches is not enough and further gain is obtained

when considering the context of these patches as shown subsequently.


A. Context-Aware CCA

For each patch ui, we define an anisotropic (typed) neighborhood system {Nc(i)}8c=1 which corresponds to the

eight spatial neighbors of ui in a regular grid; for instance when c = 1, N1(i) corresponds to the top-left neighbor

of ui (see Fig. 2). Using {Nc(.)}8c=1, we consider for each c an intrinsic adjacency matrix Wcu whose (i, k)th entry

is set as Wcu,i,k ∝ 1{k∈Nc(i)}; here 1{} is the indicator function equal to 1 iff i) the patch uk is neighbor to ui

and ii) its relative position is typed as c (c = 1 for top-left, c = 2 for left, etc. following an anticlockwise rotation),

and 0 otherwise. Similarly, we define the matrices {Wcv}c for patches {vi}i.

We introduce our main contribution; a novel context-aware approach for CCA. Considering a small subset

{(ui,vi)}i ⊂ I with known labels {yi}i, and now Pu, Pv as transformation matrices, we propose to find the

latter as

maxPu,Pv

tr(U′PuP′vVD) + β

8∑c=1

tr(U′PuP

′vVWc

vV′PvP

′uUWc′

u

)s.t. P′

uCuuPu = Id and P′vCvvPv = Id.

(1)

The non-matrix forms of these two terms are given subsequently. In this constrained maximization problem, β ≥ 0,

tr is the trace, U, V are two matrices of aligned patch pairs {(ui,vi)}i, Id is the d× d identity matrix and D is a

diagonal matrix with its given entry Dii set proportional to (a) −1 if yi = +1, (b) +1 if yi = −1 and (c) 0 if yi

is unknown. This particular setting of D makes CCA discriminant; indeed, the left-hand side term of this objective

function (equivalent to∑i,j〈P′uui,P′vvj〉Dij) aims to maximize the correlation between patch pairs with negative

labels (i.e., “no-changes”), while at the same time, it minimizes the correlation between patch pairs with positive

labels (i.e., relevant “changes”). Hence, this left-hand side term is strictly equivalent to discriminant CCA. With this

discriminative setting, the learned transformations Pu, Pv generate latent data representations φu(ui) = P′uψf (ui),

φv(vi) = P′vψf (vi), which are robust against irrelevant changes3 (i.e., ‖φu(ui) − φv(vi)‖2 0 for yi = −1)

while also being sensitive to relevant changes (i.e., ‖φu(ui)− φv(vi)‖2 is large for yi = +1). This results into a

better discrimination between “changes” and “no-changes”.

Using the above definition of {Wcu}c, {Wc

v}c, the right-hand side term of the objective function (1) is equivalent

to

β∑c

∑i,j〈P′uui,P′vvj〉

∑k,`〈P′uuk,P′vv`〉Wc

u,i,kWcv,j,`; the latter corresponds to a neighborhood (or context)

criterion which considers that a high value of the correlation 〈P′uui,P′vvj〉, in the learned latent space, should

imply high correlation values in the neighborhoods {Nc(i)×Nc(j)}c (see Fig. 2). Hence, this term enhances the

robustness of the correlation between patch pairs in the learned latent space. Put differently, if a given patch pair

(ui,vj) is surrounded by highly correlated pairs, then the correlation between (ui,vj) should be maximized and

vice-versa.

3with ψf (ui) being a feature vector associated to ui, see experiments.


B. Optimization

Considering Lagrange multipliers for the equality constraints in Eq. (1), one may show that optimality conditions

(related to the gradient of Eq. (1) w.r.t Pu, Pv and the Lagrange multipliers) lead to the generalized eigenproblem

K′vuC−1vv KvuPu = CuuPuΛ

2

with Pv = C−1vv KvuPuΛ−1,

(2)

here Kvu = VDU′ +β∑c VWc

vV′PvP

′uUWc′

u U′

+β∑c VWc′

v V′PvP′uUWc

uU′,

(3)

and Λ2 is the hadamard power of the diagonal matrix Λ. As K′vuC−1vv Kvu is Hermitian and Cuu is positive

semi-definite, the eigenproblem in Eq. (2) admits real eigenvalues (in Λ2) and its eigenvectors, when multiplied by

C12uu (resp. C

12vv), are mutually orthogonal and from an orthogonal basis; here C

12uu, C

12vv are square roots of Cuu,

Cvv respectively.

We solve the above eigenproblem iteratively. For each iteration τ , we fix P(τ)u , P

(τ)v (in Kvu) and we find the

subsequent P(τ+1)u , P

(τ+1)v by solving the eigenproblem in Eq. (2); initially, P

(0)u , P

(0)v are set using projection

matrices of standard CCA and this turns out to provide better behavior compared to random initializations (i.e., it

rapidly converges to a unique solution of objective function 1). In practice, convergence to a fixed point is observed

in less than five iterations4.

IV. EXPERIMENTAL VALIDATION

We evaluate the performances of our interactive change detection algorithm using a dataset of 4, 400 non

overlapping patch pairs (of 30×30 pixels in RGB) taken from two registered (reference and test) GeoEye-1 satellite

images of 2, 400 × 1, 652 pixels with a spatial resolution of 1.65m/pixel. These images correspond to the same

area of Jefferson (Alabama) taken respectively in 2010 and in 2011 with many changes due to tornados (building

destruction, etc.) and no-changes (including irrelevant ones as clouds). The underlying ground truth contains 4, 275

negative patch pairs (“no-changes” and irrelevant ones) and only 125 positive patch pairs (relevant changes), so

< 3% of these patches correspond to relevant changes (see Fig. 3a–c). Note that the spatial distribution of changes

in this set is globally smooth, so this dataset is suitable to evaluate the impact of our proposed context-aware CCA

method (see Sections IV-A, IV-B).

Each patch (in reference and test images) is encoded with d = 200 coefficients corresponding to its projection on

the d principal axes of PCA. These principal axes of PCA were estimated using all patches of the reference image

and capture more than 95% of the statistical variance of the data. Afterwards, each patch pair xi = (ui,vi) in I is

described either without CCA as i) ψf (vi) − ψf (ui) with ψf (ui) being the projection of ui using PCA or as ii)

φv(ψf (vi))−φu(ψf (ui)) when the CCA latent representations φu(.), φv(.) are considered; since the learned CCA

transformations are linear, the dimensionality of φu(.), φv(.) is bounded by d and in practice, we keep all the d

dimensions in order to avoid any substantial loss of information w.r.t the original patches. Performances are reported

4The complexity of solving Eq. (2) depends mainly on the dimension d (d� n).


(a) (b) (c)

Fig. 3. (a–c) Area taken from the reference image (before tornados) and test image (after tornados) as well as their ground truth in “Jefferson”;

“red” stands for relevant changes while “white” stands for irrelevant changes (mainly clouds), the remaining locations correspond to no-changes.

using equal error rate (EER) on unlabeled data of I. EER is the balanced generalization error that equally weights

errors in “change”, “no-change” classes and it is defined as 12 (

# of incorrect change detections# of change detections + # of undetected actual changes

# of actual changes )×100.

Smaller EER implies better performance.

A. Impact of Relevance Feedback

We compare our RF-based change detection criteria {ft(.)}Tt=0, against four RF-free criteria; i.e., independent

of t

(i) Image difference: a pair xi = (ui,vi) ∈ I is declared as a change iff ‖ψf (ui)− ψf (vi)‖22 0.

(ii) Large scale (monolithic) SVM: we train an SVM decision function (denoted f ) and we use it to detect changes

in I. The training set of f (denoted T ′ = {(x′i,y′i)}i) is different from I; it includes 220 positive examples and

4, 180 negative examples extracted from other GeoEye-1 satellite images (including 2, 400 × 1, 652 pixels) of the

nearby area of Tuscaloosa (Alabama) taken respectively in 2010 and in 2011 with many changes due to tornados

that also happen in 2011. Note that the set T ′ is much larger than the one used to train the final classifier fT−1

(which has only∑T−1t=0 |Dt| = 160 examples at the end of our relevance feedback process).

(iii) Large scale (monolithic) CA-DCCA5: this setting is similar to monolithic SVM, but now SVM is trained on

top of CA-DCCA features instead of PCA.

(iv) Large scale (monolithic) FDA6: this setting is also similar to monolithic SVM with the only difference being

the used criterion; FDA is used as a classifier instead of SVM.

Fig. 4a shows EERs of these baselines (i)–(iv) as well as our proposed RF method (with and without CA-DCCA);

we first observe that monolithic CA-DCCA achieves the best performances among all the baselines, and this clearly

shows that the CA-DCCA learned on the area of “Tuscaloosa” is also (relatively) suitable for “Jefferson”. Whereas

these baselines and our proposed RF method (for t ≤ 1) have high EERs, our RF method rapidly reduces the

EER and overtakes all the monolithic criteria, at the end of the iterative process. This comes essentially from the

5CA-DCCA stands for context-aware discriminant CCA transformations learned here using patch pairs belonging to the area of “Tuscaloosa”.

Once trained, all these transformations and SVM are applied to patches of “Jefferson”.6Fisher Discriminant Analysis.


1 3 5 7 9

10

15

20

25

30

35

Iteration number (t)

Equal E

rror

Rate

(%

)

Image Difference

Large Scale (Monolithic) FDA

Large Scale (Monolithic) SVM

Large Scale (Monolithic) CA−DCCA

Proposed RF with no CCA

Proposed RF with CA−DCCA

1 3 5 7 910

15

25

35

45

Iteration number (t)

Equal E

rror

Rate

(%

)

no CCAStandard CCADiscriminant CCAContext−Aware Standard CCAContext−Aware Discriminant CCAFDA

−8 −6 −4 −2 0 211

12

13

14

15

16

17

18

19

log2(β)

Equal E

rror

Rate

(%

)

Context−Aware Discriminant CCA

(a) (b) (c)

Fig. 4. (a) Comparison of our RF method against several baselines: image difference and large scale (monolithic) criteria; in this figure, noCCA

means that the underlying SVMs are learned on top of PCA features. (b) Evolution of EER results w.r.t the iteration number (t), for SVMs

built on top of different CCA versions (for context-aware CCAs, β is set to 0.15) and also using FDA; again noCCA means that SVMs are

learned on top of PCA features. (c) This figure shows the evolution of EER (of the RF-based SVMs built on top of context-aware discriminant

CCA) w.r.t β; these values correspond to EER performances at the end of the RF process (i.e., when t = 9). Note that β → 0 is equivalent to

discriminant CCA. All these results are obtained by averaging EERs of 10 RF runs.

adaptation of decision functions {ft(.)}t to the user’s intention as well as to reference and test images in I. This

is also due to the positive impact of CA-DCCA on RF as shown in Fig. 4 and also subsequently.

B. Impact of Different CCA

Now we study the impact of different CCA methods on the RF performances. Fig. 4b shows the RF performances

without CCA and with (standard, discriminant and context-aware) CCA w.r.t the iteration number t; in these

experiments, we consider the context-aware version of both standard and discriminant CCA. We also compare the

performance of all these CCA versions against FDA; the setting of the latter is similar to all the CCA versions

with the only difference being the use of FDA instead of CCA+SVM (as also discussed in Section IV-A). Note

that all these results were obtained by averaging EERs of 10 RF runs, each one corresponds to a random setting

of display D0 (see again Section II). These EERs decrease as t increases and reach their smallest values at the

end of the iterative process, i.e., when decision criteria {ft(.)}t are well trained/adapted to the reference and test

satellite images, and this happens after 10 iterations only. Fig. 4c shows the EERs of our RF algorithm (after

10 iterations) built upon context-aware discriminant CCA; these EERs globally decrease as β increases/reaches

intermediate values and EERs increase again for larger values of β. From all these observations, it is clear that the

proposed context-aware CCA combined with RF has a clear gain compared to the other settings and it is able to

find relevant changes and discard many irrelevant ones (in few iterations).

Finally, CCA and SVM learning/classification (as well as the display selection strategy discussed in section II) run

promptly (in less than 1s at each iteration t on a Quad Core 3.60GHz PC), provided that PCA features are extracted

off-line; so change detection results are iteratively updated in real-time.


V. CONCLUSION

We introduced in this paper a novel change detection algorithm based on relevance feedback and a new variant of

CCA referred to as context-aware CCA. Our method considers both labeled and unlabeled data when learning the

CCA transformations. This is achieved by optimizing an objective function mixing two terms: the first one relies on

a discriminative setting that maximizes (resp. minimizes) correlations between “no changes” (resp. “changes”). The

second term acts as a regularizer that makes correlations spatially smooth and provides us with robust context-aware

latent representations. As shown through experiments, the relative gain of our method is substantial especially when

changes are smoothly spread; this is valuable for changes impacted by natural hazards following regular paths such

as tornados.

REFERENCES

[1] K. Rokni, A. Ahmad, K. Solaimani, and S. Hazini, “A new approach for surface water change detection: integration of pixel level image

fusion and image classification techniques,” International Journal of Applied Earth Observation and Geoinformation, vol. 34, pp. 226–234,

2015.

[2] S. Jamali, P. Jonsson, L. Eklundh, J. Ardo, and J. Seaquist, “Detecting changes in vegetation trends using time series segmentation,”

Remote Sensing of Environment, vol. 156, pp. 182–195, 2015.

[3] Z. Zhu and C. E. Woodcock, “Continuous change detection and classification of land cover using all available landsat data,” Remote

sensing of Environment, vol. 144, pp. 152–171, 2014.

[4] D. Brunner, G. Lemoine, and L. Bruzzone, “Earthquake damage assessment of buildings using vhr optical and sar imagery,” IEEE Trans.

Geosc. Remote Sens., vol. 48, no. 5, pp. 2403–2420, 2010.

[5] H. Gokon, J. Post, E. Stein, S. Martinis, A. Twele, M. Muck, C. Geiss, S. Koshimura, and M. Matsuoka, “A method for detecting buildings

destroyed by the 2011 tohoku earthquake and tsunami using multitemporal terrasar-x data,” GRSL, vol. 12, no. 6, pp. 1277–1281, 2015.

[6] N. Longbotham, F. Pacifici, T. Glenn, A. Zare, M. Volpi, D. Tuia, E. Christophe, J. Michel, J. Inglada, J. Chanussot et al., “Multi-modal

change detection, application to the detection of flooded areas: Outcome of the 2009–2010 data fusion contest,” Selected Topics in Applied

Earth Observations and Remote Sensing, vol. 5, no. 1, pp. 331–342, 2012.

[7] J. Deng, K. Wang, Y. Deng, and G. Qi, “Pca-based land-use change detection and analysis using multitemporal and multisensor satellite

data,” IJRS, vol. 29, no. 16, pp. 4823–4838, 2008.

[8] E. F. Lambin and A. H. Strahlers, “Change-vector analysis in multitemporal space: a tool to detect and categorize land-cover change

processes using high temporal-resolution satellite data,” Remote Sensing of Environment, vol. 48, no. 2, pp. 231–244, 1994.

[9] R. Radke, S. Andra, O. Al-Kofahi, and B. Roysam, “Image change detection algorithms: A systematic survey,” IEEE Trans. on Im Proc,

vol. 14, no. 3, pp. 294–307, 2005.

[10] G. Moser and S. Serpico, “Generalized minimum error thresholding for unsupervised change detection from amplitude sar imagery,” IEEE

TGRS, vol. 44, no. 10, pp. 2972–2983, 2006.

[11] S. Liu, L. Bruzzone, F. Bovolo, M. Zanetti, and P. Du, “Sequential spectral change vector analysis for iteratively discovering and detecting

multiple changes in hyperspectral images,” TGRS, vol. 53, no. 8, pp. 4363–4378, 2015.

[12] M. Hussain, D. Chen, A. Cheng, H. Wei, and D. Stanley, “Change detection from remotely sensed images: From pixel-based to object-based

approaches,” JPRS, vol. 80, pp. 91–106, 2013.

[13] G. Chen, G. J. Hay, L. M. Carvalho, and M. A. Wulder, “Object-based change detection,” IJRS, vol. 33, no. 14, pp. 4434–4457, 2012.

[14] J. Zhu, Q. Guo, D. Li, and T. C. Harmon, “Reducing mis-registration and shadow effects on change detection in wetlands,” Photogrammetric

Engineering & Remote Sensing, vol. 77, no. 4, pp. 325–334, 2011.

[15] A. Fournier, P. Weiss, L. Blanc-Fraud, and G. Aubert, “A contrast equalization procedure for change detection algorithms: applications to

remotely sensed images of urban areas,” In ICPR, 2008.

[16] N. Bourdis, D. Marraud, and H. Sahbi, “Camera pose estimation using visual servoing for aerial video change detection,” in IEEE IGARSS,

2012, pp. 3459–3462.

[17] Carlotto, “Detecting change in images with parallax,” In Society of Photo-Optical Instrumentation Engineers, 2007.


[18] N. Bourdis, D. Marraud, and H. Sahbi, “Constrained optical flow for aerial image change detection,” in IEEE IGARSS, 2011, pp. 4176–4179.

[19] X. Chen, L. Vierling, and D. Deering, “A simple and effective radiometric correction method to improve landscape change detection across

sensors and across time,” RSE, vol. 98, no. 1, pp. 63–79, 2005.

[20] S. Leprince, S. Barbot, F. Ayoub, and J.-P. Avouac, “Automatic and precise orthorectification, coregistration, and subpixel correlation of

satellite images, application to ground deformation measurements,” TGRS, vol. 45, no. 6, pp. 1529–1558, 2007.

[21] Pollard, “Comprehensive 3d change detection using volumetric appearance modeling,” Phd, Brown University, 2009.

[22] R. Wiemker, “An iterative spectral-spatial bayesian labeling approach for unsupervised robust change detection on remotely sensed

multispectral imagery,” In Proc. CAIP, LNCS 1296, pp. 263–270, 1997.

[23] M. Molinier, J. Laaksonen, and T. Hame, “Detecting man-made struct and changes in sat imagery with a content-based information retrieval

system built on self-organizing maps,” TGRS, vol. 45, no. 4, 2007.

[24] O. Sjahputera, G. Scott, B. Claywell, M. Klaric, N. Hudson, J. Keller, and C. Davis, “Clustering of detected changes in high-resolution

satellite imagery using a stabilized competitive agglomeration algorithm,” IEEE TGRS, vol. 49, no. 12, 2011.

[25] A. A. Nielsen, “The regularized iteratively reweighted mad method for change detection in multi-and hyperspectral data,” IEEE Transactions

on Image processing, vol. 16, no. 2, pp. 463–478, 2007.

[26] C. Wu, B. Du, and L. Zhang, “Slow feature analysis for change detection in multispectral imagery,” TGRS, vol. 52, no. 5, pp. 2858–2874,

2014.

[27] N. Bourdis, D. Marraud, and H. Sahbi, “Spatio-temporal interaction for aerial video change detection,” in IGARSS, 2012, pp. 2253–2256.

[28] J. Im, J. Jensen, and J. Tullis, “Object-based change detection using correlation image analysis and image segmentation,” International

Journal of Remote Sensing, vol. 29, no. 2, pp. 399–423, 2008.

[29] C. Tarantino, M. Adamo, R. Lucas, and P. Blonda, “Detection of changes in semi-natural grasslands by cross correlation analysis with

worldview-2 images and new landsat 8 data,” RSE, vol. 175, pp. 65–72, 2016.

[30] H. Sahbi, “Discriminant canonial correlation analysis for interactive satellite image change detection,” IEEE IGARSS, 2015.

[31] F. Fleuret and H. Sahbi, “Scale invariance of support vector machines based on the triangular kernel,” SCTV (part of ICCV), 2003.

[32] M. Ferecatu, “Image retrieval with active relevance feedback using both visual and keyword-based descriptors,” PhD thesis, 2005.

[33] H. Sahbi, “Relevance feedback for satellite image change detection,” in IEEE ICASSP, 2013, pp. 1503–1507.

[34] D. Hardoon, S. Szedmak, and J. Shawe-Taylor, “Canonical correlation analysis: An overview with application to learning methods,” Neural

computation, vol. 16, no. 12, pp. 2639–2664, 2004.

Documents

IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1 Interactive …sahbi/grsl2017.pdf · 2017-09-21 · IEEE GEOSCIENCE AND REMOTE SENSING LETTERS 1 Interactive Satellite Image Change Detection