Remote Sensing Image Representation based on

Hierarchical Histogram Propagation

Jefersson Ale Dos Santos, Otavio Penatti, Ricardo Da Silva Torres,

Philippe-Henri Gosselin, Sylvie Philipp-Foliguet, Alexandre Xavier Falcao

To cite this version:

Jefersson Ale Dos Santos, Otavio Penatti, Ricardo Da Silva Torres, Philippe-Henri Gosselin,Sylvie Philipp-Foliguet, et al.. Remote Sensing Image Representation based on HierarchicalHistogram Propagation. IEEE International Geoscience and Remote Sensing Symposium, Jul2013, Melbourne, Australia. pp.4, 2013. <hal-00861379>

HAL Id: hal-00861379

https://hal.archives-ouvertes.fr/hal-00861379

Submitted on 12 Sep 2013

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, estdestinee au depot et a la diffusion de documentsscientifiques de niveau recherche, publies ou non,emanant des etablissements d’enseignement et derecherche francais ou etrangers, des laboratoirespublics ou prives.

REMOTE SENSING IMAGE REPRESENTATION BASED ONHIERARCHICAL HISTOGRAM PROPAGATION

Jefersson A. dos Santos1,2, Otavio A. B. Penatti1, Ricardo da S. Torres1,Philippe-H. Gosselin2, Sylvie Philipp-Foliguet2, Alexandre X. Falcao1

1RECOD Lab – Institute of Computing, University of Campinas, Brazil2ETIS, CNRS, ENSEA, University of Cergy-Pontoise, France

ABSTRACTMany methods have been recently proposed to dealwith the large amount of data provided by high-resolution remote sensing technologies. Several ofthese methods rely on the use of image segmentationalgorithms for delineating target objects. However, acommon issue in geographic object-based applicationsis the definition of the appropriate data representationscale, a problem that can be addressed by exploitingmultiscale segmentation. The use of multiple scales,however, raises new challenges related to the definitionof effective and efficient mechanisms for extractingfeatures. In this paper, we address the problem ofextracting histogram-based features from a hierarchyof regions for multiscale classification. The strategy,called H-Propagation, exploits the existing relation-ships among regions in a hierarchy to iteratively prop-agate features along multiple scales. The proposedmethod speeds up the feature extraction process andyields good results when compared with global low-level extraction approaches.

Index Terms— Histogram propagation, image rep-resentation, multiscale analysis, GEOBIA.

1. INTRODUCTION

Remote sensing images are often used as data sourcefor land cover studies in many applications [1]. A com-mon problem in these applications is the definition ofthe data representation scale (the size of the segmentedregions or block of pixels) [2, 3]. To address the seg-mentation scale problem, several approaches have beenproposed considering multiscale analysis for applica-tions that handle remote sensing images [1, 2, 4–6]. Inthese approaches, the feature extraction at various seg-mentation scales is an essential step. However, depend-ing on the strategy, the extraction can be a very costlyprocess. If we apply the same feature extraction algo-rithm for all regions of different segmentation scales,

This work was supported by CAPES (592/08 and BEX5224101), FAPESP (Grants 2008/58528-2, 2010/52113-5,2012/16253-2, and 2012/18768-0), and CNPq (303673/2010-9,484254/2012-0, and 306580/2012-8).

for example, the pixels in the image would need to beaccessed at least once for each scale.

Aiming at addressing this issue, we have recentlyproposed an strategy called BoW-Propagation [7],which exploits the bag-of-word concept [8, 9] to iter-atively propagate texture features along the hierarchyfrom the finest regions to the coarsest ones. That strat-egy speeds up the extraction process by avoiding thecomputation of low-level features from all regions ofthe hierarchy.

In this paper, we propose an approach, called H-Propagation to extract histogram-based features from ahierarchy of segmented regions. This approach, whichis an extension of the BoW-Propagation, relies on pro-cessing only the image pixels in the base of the hierar-chy (the finest region scale). The features are quicklypropagated to the upper scales by exploiting the hier-archical association among regions at different scales.The features are then propagated to the other scales. Atthe end, all regions in the hierarchy are represented bya histogram.

2. BOW-PROPAGATION

The BoW-Propagation [7], which is based on the bag-of-visual-word model [10], is designed to extract fea-tures from a hierarchy of segmented regions [11]. Itrelies on processing only the image pixels in the baseof the hierarchy (the finest regions scale). It exploitsthe hierarchical relationship among regions at differentscales to compute the features.

The strategy starts by creating a visual dictionarybased on low-level features extracted from the pixellevel (the base of the hierarchy), as shown in Figure 1.The low-level feature space is quantized, creating thevisual words, and each region in the base of the hier-archy is described according to that dictionary. Thefeatures are then propagated to the other scales. At theend, all regions in the hierarchy will be represented bya bag of visual words.

Figure 2 illustrates a schema to represent a seg-mented region by using dense sampling through a bagof words. The low-level features extracted from the

Visual Dictionary

VisualWordsQuantizationFeatures

w1w2

w3

w4

w5

w6

w1 w2 w6, ,...,

Fig. 1. Construction of a visual dictionary to describe a remote sensing image. The features are extracted fromgroups of pixels (e.g., tiles or segmented regions), the feature space is quantized so that each cluster correspondsto a visual word wi.

w1w2

w3

w4

w5

w6

w1 w2 w3 w4 w5 w6

Assignment of Visual Words

...

Assignment Vectors

Pooling

=

Bag of Visual Words

Fig. 2. Representation of a segmented region based on a visual dictionary with dense sampling feature extraction.

internal points are assigned to visual words and com-bined by a pooling function.

3. THE PROPOSED HISTOGRAM FEATUREPROPAGATION

The histogram propagation (H-Propagation) consistsin estimating the feature histogram representation ofa region R, given the low-level histograms extractedfrom the R subregions Γ(R). It is an extension of theBoW-propagation in the sense it propagates any kind oflow-level features based on histograms from fine scalesto the coarsest ones.

Let Pλxand Pλy

be partitions obtained from thehierarchy H at the scales λx e λy , respectively. Weconsider that Pλx > Pλb

, i.e, Pλx is coarser than Pλy .Let R ∈ Pλx be a region from the partition Pλx . Wecall subregion of R the region R ∈ Pλy

such that R ⊆R.

The set Γ(R), which is composed of the subregionsof R in the partition Pλy , is given by Γ(R) = {∀R ∈Pλy|p ∈ R ∩ p ∈ R}, where p is a pixel. The set of

subregions of R in a finer scale are all the regions Rthat have all pixels inside R and inside R.

The principle of H-propagation is to compute thefeature histogram hR, which describes region R, bycombining the histograms of subregions Γ(R):

hR = f{hRc| Rc ∈ Γ(R)} (1)

where f is a combination function.

Algorithm 1 presents the proposed H-propagation.The first step is to extract low-level features from theregions in the finest scale λ1 (line 1). The “propaga-tion loop” is responsible for propagating the featuresto other scales (lines 2 to 6). For all regions R from apartition Pλx , the histogram hR is computed based onthe Γ(R) histograms, which is defined by Equation 1(line 4).

Algorithm 1 H-Propagation

1 Extract low-level feature histograms from the regionsin the finest scale λ1

2 For i← 2 to n do3 For all R ∈ Pλi

do4 Compute the histogram hR based on the

Γ(R) histograms5 End for6 End for

Figure 3 illustrates an example of the use of thecombination function f to compute the histogram hrof a region r. The region r ∈ Pλ2 is composed of theset of subregions Γ(r) = {a, b, c} at the scale λ1. Fig-ure 3 (a) illustrates, in gray, the region r and its subre-gions Γ(r) in the hierarchy of regions. In Figure 3 (b),the histogram hr is computed based on the function f :hr = f(ha, hb, hc). Figure 4 illustrates the computa-tion of hr by using the max operator as combinationfunction.

H-propagation does not quantize the low-level fea-

3

2

1

r

a b c

a

bc

r

(a) (b)

Fig. 3. Computing the histogram hr of region r by combining the histrograms ha, hb, and hc from the subregionsa, b, and c.

ture space to create a visual dictionary. Another dif-ference, when compared with the BoW-propagation, isthat H-propagation propagates histogram bins insteadof the probabilities of visual words. BoW-propagationis suitable for propagating low-level local features. H-propagation, in turn, is designed only for global de-scriptors based on histogram representations.

An important issue is the definition of the propa-gation function f in the case of low-level histograms.Contrarily to the propagation of visual words, we usethe average function instead of the max function. It isexpected that with the average propagation, the qualityof the histograms will be the same as that performedby the extraction directly from the pixels at all scalesof the hierarchy.

a

bc

maxpropagation

scale

scale

i

i-1

r

0.2 0.2 0.3 0.2 0.1 0.0 0.2 0.1

0.1 0.0 0.3 0.1 0.1 0.0 0.2 0.0

0.1 0.2 0.1 0.1 0.0 0.0 0.2 0.0

0.2 0.1 0.1 0.2 0.1 0.0 0.1 0.1

Fig. 4. Feature propagation example using a max pool-ing operation.

4. SIMULATIONS

We have used two different datasets in our experi-ments. We refer to the dataset according to the targetregions: COFFEE and URBAN. We carried out ex-periments with ten different combinations of the nine

subimages for each dataset (three for training, three forvalidation, and three for testing). The COFFEE datasetis a SPOT image with 2.3m of spatial resolution fromMonte Santo de Minas, Brazil. The URBAN datasetis a Quickbird image with 0.6m of spatial resolutionfrom Campinas, Brazil. More details about the datasetscan be found in [1]. We define the scales according tothe principle of dichotomic cuts proposed in [11]. Thehigher the index, the coarser is the scale. The scale λ0is the pixel level. We used linear SVMs to obtain theclassification results. It is evaluated by computing theoverall accuracy, kappa index, and tau index for theclassified images [12].

We have tested the proposed approaches for colorfeature propagation. We have selected BIC descrip-tor [13] since it produced good results in previousworks considering multiscale classification [4, 14, 15].We compare the propagation approaches against BIClow-level feature extraction [13].

BIC BoW-Propagation was computed by using:max pooling function, dictionary size of 103 words,and soft assignment (σ = 0.1). We have extracted low-level features from a dense sampling by overlappingsquares with 4× 4 pixels. BIC H-Propagation, in turn,was computed by using the avg pooling function.

Table 1 presents the classification using the BIC de-scriptor with BoW-Propagation, Histogram Propaga-tion, and low-level extraction (Global Descriptor) forthe COFFEE dataset.

Concerning the results for the COFFEE dataset,H-Propagation and the Global Descriptor presentthe same overall accuracy (around 80%). The samecan be observed for kappa and tau indexes. BoW-Propagation achieved results slightly worse than theother two approaches for the three computed mea-sures. Regarding the URBAN dataset, H-Propagationand Global Descriptor obtained the same overall ac-curacy, Kappa, and Tau (∼ 70%, 0.31, and 0.47,

Table 1. Classification results for the BIC descriptor using BoW-Propagation, H-Propagation, and global featureextraction for the COFFEE and the URBAN datasets at segmentation scale λ3.

COFFEE URBANMethod O.A. (%) Kappa (κ) Tau (τ ) O.A. (%) Kappa (κ) Tau (τ )

BoW-Propagation 73.41± 2.76 0.25± 0.03 0.36± 0.02 67.03± 2.65 0.26± 0.03 0.47 ± 0.02H-Propagation 79.97 ± 1.76 0.46 ± 0.02 0.54 ± 0.02 69.86 ± 4.76 0.31 ± 0.05 0.47 ± 0.04

Global Descriptor 80.07 ± 1.81 0.47 ± 0.02 0.54 ± 0.02 69.63 ± 3.33 0.31 ± 0.04 0.47 ± 0.03

respectively). The BoW-Propagation approach yieldsslightly worse results than the other methods concern-ing overall accuray and Kappa index. The Tau indexwas the same (0.47).

5. CONCLUSION

The proposed H-propagation approach revealed besuitable for saving time on feature extraction froma hierarchy of segmented regions. Regarding colorfeatures, BOW-Propagation seems to be promising,but it requires many setup parameters. However, H-Propagation shows that it is possible to compute low-level features based only on the extracted hierarchy.Furthermore, these features can be propagated withoutlosses in terms of representation quality.

Future work considers the study the use of theH-propagation method in other applications, such ascontent-based image retrieval and multimedia foren-sics.

6. REFERENCES

[1] J. A. dos Santos, P.H. Gosselin, S. Philipp-Foliguet,R. da S. Torres, and A. X. Falcao, “Interactive mul-tiscale classification of high-resolution remote sensingimages,” Selected Topics in Applied Earth Observa-tions and Remote Sensing, IEEE Journal of, 2013, Toappear.

[2] Y. Tarabalka, J.C. Tilton, J.A. Benediktsson, andJ. Chanussot, “A marker-based approach for the auto-mated selection of a single segmentation from a hierar-chical set of image segmentations,” Selected Topics inApplied Earth Observations and Remote Sensing, IEEEJournal of, vol. 5, no. 1, pp. 262 –272, feb. 2012.

[3] J. A. dos Santos, C. D. Ferreira, R. da S.Torres, M. A.Goncalves, and R. A. C. Lamparelli, “A relevance feed-back method based on genetic programming for classi-fication of remote sensing images,” Information Sci-ences, vol. 181, no. 13, pp. 2671–2684, 2011.

[4] J. A. dos Santos, P.H. Gosselin, S. Philipp-Foliguet,R. da S. Torres, and A. X. Falcao, “Multiscale clas-sification of remote sensing images,” Geoscience andRemote Sensing, IEEE Transactions on, vol. 50, no. 10,pp. 3764–3775, 2012.

[5] J. Munoz-Mari, D. Tuia, and G. Camps-Valls, “Semisu-pervised classification of remote sensing images withactive queries,” Geoscience and Remote Sensing, IEEE

Transactions on, vol. 50, no. 10, pp. 3751 –3763, oct.2012.

[6] A. Alonso-Gonzalez, S. Valero, J. Chanussot,C. Lopez-Martınez, and P. Salembier, “Process-ing multidimensional sar and hyperspectral imageswith binary partition tree,” Proceedings of the IEEE,vol. 101, no. 3, pp. 723–747, 2013.

[7] J. A. dos Santos, O. A. B. Penatti, R. da S. Torres,P-H. Gosselin, S. Philipp-Foliguet, and A. X. Falcao,“Improving texture description in remote sensing imagemulti-scale classification tasks by using visual words,”in International Conference on Pattern Recognition,Tsukuba, Japan, November 2012, pp. 3090–3093.

[8] Y.-L. Boureau, F. Bach, Y. LeCun, and J. Ponce,“Learning mid-level features for recognition,” Confer-ence on Computer Vision and Pattern Recognition, pp.2559–2566, 2010.

[9] J. C. van Gemert, C. J. Veenman, A. W. M. Smeulders,and J-M Geusebroek, “Visual word ambiguity,” Trans-actions on Pattern Analysis and Machine Intelligence,vol. 32, pp. 1271–1283, 2010.

[10] J. Sivic and A. Zisserman, “Video google: a text re-trieval approach to object matching in videos,” in In-ternational Conference on Computer Vision, 2003, pp.1470–1477 vol.2.

[11] L. Guigues, J. Cocquerez, and H. Le Men, “Scale-setsimage analysis,” International Journal of Computer Vi-sion, vol. 68, pp. 289–317, 2006.

[12] Zhenkui Ma and Roland Redmond, “Tau coefficientsfor accuracy assessment of classification of remotesensing data,” Photogrammetric Engineering and Re-mote Sensing, vol. 61, no. 4, pp. 439–453, 1995.

[13] R. de O. Stehling, M. A. Nascimento, and A. X.Falcao, “A compact and efficient image retrieval ap-proach based on border/interior pixel classification,” inCIKM, New York, NY, USA, 2002, pp. 102–109.

[14] J. A. dos Santos, F. A. Faria, R. da S. Torres, A. Rocha,P-H. Gosselin, S. Philipp-Foliguet, and A. X. Falcao,“Descriptor correlation analysis for remote sensing im-age multi-scale classification,” in International Confer-ence on Pattern Recognition, Tsukuba, Japan, Novem-ber 2012, pp. 3078–3081.

[15] J. A. dos Santos, O. A. B. Penatti, and R. da S. Torres,“Evaluating the potential of texture and color descrip-tors for remote sensing image retrieval and classifica-tion,” in The International Conference on ComputerVision Theory and Applications, Angers, France, May2010, pp. 203–208.