OBJECT-ORIENTED APPROACH AND TEXTURE ANALYSIS FOR CHANGE DETECTION IN VERY HIGH RESOLUTION IMAGES

Antoine Lefebvre, Thomas Corpetti and Laurence Hubert-Moy

COSTEL, UMR CNRS 6554 LETG – IFR 90 CAREN

Université Rennes 2 Place du recteur Henri Le Moal, CS 24307

35043 Rennes Cedex - France

ABSTRACT The objective of this paper is to develop an object-based change detection method able to qualify the nature of changes in landscapes from remotely sensed images, in terms of geometry and content. The originality of the approach consists in jointly dealing with the analysis of the object contours and the analysis of texture evolution. The method is applied on grassy strips, which are landscape buffers between crops and hydrologic networks. A geometric change index that quantify the intensity of change and that qualify it according its properties is proposed. We also present a content change index that discriminates partial change from diffuse change that affect the object of interest. It turned out that the geometric changes, most of abrupt content changes and some of subtle changes have been accurately detected. Lastly, our approach is suitable on airborne data with a Very High Resolution data and can be generallized to spaceborne images.

Index Terms— Object oriented methods, Remote sensing, Wavelet transforms, Image texture analysis, Agriculture

1. INTRODUCTION In the remote sensing community, most of change detection methods have been developed to detect abrupt changes from low or medium resolution imagery [1;2]. These methods, often based on a pixel analysis, are limited in presence of Very High Resolution images (VHR). Due to the amount of data and to the spatial relationship of the pixels, their benefit is indeed reduced since it does not exploit any spatial information.

In the last five years, some object-oriented approaches have successfully been applied to VHR images [3;4]. These techniques allow to identify landscape features and to extract some information about shape, content and context from a single image [5]. Nevertheless, at the moment, the problem of change detection based on such objects approaches has rarely been studied. Most of techniques rely indeed on pixel intensities but are not focused on additional dimensions such as texture or shape. The objective of this study is to develop an object-based change detection method able to qualify the nature of changes in landscapes from remotely sensed images, in terms of geometry and content.

2. CHANGE DETECTION

Two types of changes are generally distinguished in land-cover: conversion and modification. The conversion represents a change from one cover type to another. The modification corresponds to a change of condition within a land-cover category [6]. In the one hand, the conversion can be considered like an abrupt change as the complete replacement of one cover type by another. On the other hand, the modification is more subtle since this can affect the character of the land-cover without changing its classification. According to Coppin et al. [7], one of the main challenges facing ecosystem change monitoring is to detect modifications in addition to conversions.

The change detection can be evaluated in a spatial context. The knowledge on the spatial extent of change provides information that help to identify the nature of change and its driving factors. A typology of spatial changes has been proposed by Khorram et al. [8], which highlights fragmentation, size and shape modification as well as location shifting. Some recent image segmentation applications identify geometric and content changes in multitemporal object-based image analysis [5].

It is then of primary importance to develop some object-oriented approaches that consider any conversion and modification which may occur (see for instance Figure 1). More precisely, both partial and diffuse changes, as well as abrupt and subtle changes within objects have to be considered. This implies to extend the focus on the texture of objects instead of relying only on individual pixel intensities or spectral signatures. The originality of this approach consists in jointly dealing with the analysis of the object contours and the analysis of texture evolution.

3. METHOD

The proposed approach can be divided into three steps: 1-segmentation, 2-geometric change detection and 3-content change detection. 3.1. Segmentation We chose to extract the objects of interest with an object-oriented approach. It is actually known that such approaches give satisfactory results when applied on Very High Resolution images [9;10]. Objects with common properties, such as color (spectral), texture and shape, are segmented into homogenous regions. In this

IV - 663978-1-4244-2808-3/08/$25.00 ©2008 IEEE IGARSS 2008

Original image Original image

Original image

Image with subtlediffuse change

Image with abruptdiffuse change

Image with subtlepartial change

Image with abruptpartial change

Original image

Abrupt change (conversion)Subtle change (modification)

Figure 1. A typology of object content changes

study, a multiresolution-based image segmentation is used to define different segmentation scales, that enables to build up a hierarchical network of image objects prior to classify them in different change categories. 3.2. Geometric change detection The geometric change detection is based on the analysis of the object contours. Two change indices that detect existence or non existence, size and shape as well as location changes that may occur are defined. Many shape indices can be computed from image objects. In this study, the variation of object surface as object length and width are estimated. This allows us to assess and interpret object expansion or contraction. Lastly, comparing both the coordinates of the object centroïds and the object min-max coordinates from one date to another assesses the location shifting. 3.3. Content change detection As already mentioned, the content change detection does not rely on pixel intensities but is rather focused on their spatial arrangement. We chose to perform such analysis by comparing the changes in terms of texture. To that end, a wavelet transform has been chosen [11] since it is well known that such approaches are able to discriminate the textures [12]. 3.3.1. Quantifying the changes The proposed methodology consists in two steps : 1) from an original image, four sub-bands corresponding to the wavelet decomposition are generated: an approximation of the original image, the horizontal details, the vertical details and the diagonal details. In this study, a second order wavelet packet is used: the

four sub-bands are decomposed in turn, and 16 sub-bands are then obtained; 2) It has been shown that the distribution of the coefficients included in each sub-bands can be modeled using a Generalized Gaussian Density (GGD) [12]. A GGD is defined by 2 coefficients, and , which are respectively the shape and scale parameter. Therefore, the content change detection is based on the analysis of these coefficients from date to date. The and coefficients are estimated from the histogram of the coefficients distribution by maximum likelihood estimation [13;14]. The similarity between each GGD pair is evaluated with the Kullback-Liebler Distance (KLD). Furthermore, an overall KLD is calculated by the sum of KLD from each GGD pair. 3.3.2. Qualifying the changes Firstly, the content change detection algorithm is applied to the overall image object. At this step, it is possible to detect diffuse and abrupt changes with a threshold from the analysis of overall KLD variation. Secondly, the image object is subdivided into homogeneous blocs and the content change detection algorithm is applied on them so that this multi-scale texture analysis discriminates a diffuse change to a partial one. To evaluate the nature of the changes, some dissimilarity indexes are used. These latter are the mean and the relative standard deviation (RSD) of overall KLD of all blocs. We assume that a partial and a subtle change may occur when the mean and the RSD is low and a partial and abrupt change may occur when the mean is high and the RSD is low. Some user-specific thresholds are computed to discriminate diffuse/partial change and abrupt/subtle one. These parameters are fixed once and are the same for a family of images. Let us outline that this approach is able to exhibit the nature as well as the location of a partial change.

IV - 664

120 meters

120 meters

120 meters

120 meters

(a) (b)

(c) (d)

year 2005 year 2006 year 2005 year 2006

year 2005 year 2006 year 2005 year 2006

Figure 2. Examples of content change applied to grassy strip: (a) a diffuse and subtle change, (b) a diffuse and abrupt change, (c) a partial and subtle change, (d) a partial and abrupt change.

4. APPLICATION TO GRASSY STRIPS In this study, the method is applied to grassy strips, that are landscape buffers between crops and hydrologic networks (rivers, ditches ...). Establishing and maintaining grass margins is one of the objectives of the European Union's Common Agricultural Policy, in order to maintain biodiversity, limit water pollution and soil erosion. The monitoring of changes in such landscape features like their implantation duration, minimum surface, or land-cover type, is very important since they can alter their functionalities. 4.1. Study site and data sets The study site is located in the north-east of Ille-et-Vilaine department in France. It is a mixed woodland and pasture-land of 9350 hectares. This site is endorsed by the french national CNRS program "Zone Atelier" and the European research network "Rex-Alter Net". The main objective of this research is to understand and evaluate the effects of agricultural activities on landscapes and biodiversity. In this context, the effect of grassy strips on different species of plants and animals is currently under evaluation.

In our application, airborne images have been acquired during three successive years (2005, 2006, 2007). Their quality is poor due to several artifacts (i.e. pitch and roll, different weather conditions), and strengthens the difficulty of change detection.

The method has been applied to 18 samples of grassy strips. A reference map has been generated from aerial photographs that have been manually-interpretated based on expert-knowledge.

4.2. Results The first results obtained for geometric changes show that resulted location shiftings are partly distorted due to the presence and the expansion of wooded hedgerows, depending on the image acquisition dates and time. Actually, a total increase of 0.17 ha is observed respectively from 2005 to 2006 and whereas, their total area is quite the same from 2006 to 2007. As general trend, results show that the width of grassy strips remain stable whereas their length is more fluctuant.

For the content change detection, the results point out that compared with a map based on reference data, 12 out of 18 changes were detected accurately. The threshold values have been chosen to correspond to the typology of changes (Figure 1). The diffuse and subtle changes correspond to a very small change in texture. For instance, the texture of a grassy strip is similar to a grassy cover at the first date (Figure 2a), whereas this is the same land cover but some characteristics such as the height of cover, the percentage of coverage are slightly different. Conversely, diffuse and abrupt changes correspond to a texture that is very different from than that of a grassy cover (Figure 2b). On the whole, the method distinguishes accurately bare or ploughed soils from grassy cover. Partial and subtle changes are more difficult to detect because the dissimilarity measure is low and therefore the multi-scale analysis is less obvious. However, even in this case, results remain satisfactory: for example, Figure 2c shows differences in density of vegetation cover located in a specific area of a grassy

IV - 665

strip. In contrast, partial and abrupt changes are easier to detect. The overall KLD measured is important and the multi-scale analysis gives very significant results. In figure 2d, a grassy strip is partially ploughed in 2005 while the cover is homogeneous in 2006.

5. CONCLUSION

In this paper, we have presented an object-based change detection method able to qualify geometric and content changes. We introduced an analysis of the object contours and an analysis of texture evolution. The developed approach has been tested on grassy strips from VHR images. The experimental results bring the efficiency of our approach. In effect, it would appear that most of effective geometric and content changes have been correctly detected and can be generalized to spaceborne images.

6. ACKNOWLEDGEMENTS

This work was carried out with the financial support of the « ANR- Agence Nationale de la Recherche - The French National Research Agency » under the « Programme Agriculture et Développement Durable », project « ANR-05-PADD-11-004, COPT »

7. REFERENCES [1] S. Le Hégarat-Mascle, R. Seltz, L. Hubert-Moy, S. Corgne, S. and N. Stach, “Performance of change detection using remotely sensed data and evidential fusion: comparison of three cases of application,” International Journal of Remote Sensing, 27, pp. 3515-3532, 2006. [2] D. Lu, P. Mausel, E. Brondizio, and E. Moran, “Change detection techniques,” International Journal of Remote Sensing, 25, pp. 2365-2407, 2004. [3] L. Hubert-Moy, K. Michell, T. Corpetti, and B. Clément, “Object-oriented mapping and analysis of wetland using SPOT 5 data,” IEEE International Geoscience and Remote Sensing Symposium, IGARSS '06, Denver, USA, 2006. [4] U. C. Benz, P. Hofmann, G. Willhauck, I. Lingenfelder, and M. Heynen, “Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information,” ISPRS Journal of Photogrammetry and Remote Sensing, 58, pp. 239-258, 2004. [5] T. Blaschke, “Towards a framework for change detection based on image objects” Göttinger Geographische Abhandlungen, 113, pp. 1-9, 2005. [6] B. Turner, and W. Meyer, “Changes in land use and land cover: a global perspective,” Cambridge Univ Press, 1994. [7] P., Coppin, I. Jonckheere, K. Nackaerts, B. Muy, and E. Lambin, “Digital change detection methods in ecosystem monitoring: a review,” International Journal of Remote Sensing, 25, pp. 1565-1596, 2004. [8] S. Khorram, G. Biging, N. Chrisman, D. R. Colby, R. G. Congalton, J. E. Dobson, R. L. Fergusson, M. Goodchild, J. R. Jensen, and T. H. Mace, “Accuracy assessment of remote sensing-

derived change detection ASPRS Monograph,” Bethesda, Md., 1999. [9] J. Schiewe, L. Tufte, and M. Ehlers, “Potential and problems of multi-scale segmentation methods in remote sensing,” GeoBIT/GIS, 6, pp. 34-39, 2001. [10] M. Baatz and A. Schäpe, “Multiresolution segmentation—an optimization approach for high quality multi-scale image segmentation”. In: J. Strobl, T. Blaschke and G. Griesebner, Editors, Angewandte Geographische Informations-Verarbeitung XII, Wichmann Verlag, Karlsruhe, pp. 12–23, 2000. [11] S. A. Mallat, “Wavelet Tour of Signal Processing,” Academic Press, 1999. [12] M. Unser, “Texture Classification and Segmentation Using Wavelet Frames,” IEEE Transactions on Image Processing, 4, pp. 1549-1560, 1995. [13] M. N. Do, and M. Vetterli, “Wavelet-based texture retrieval using generalized Gaussian density and Kullback-Leibler distance,” IEEE Transactions on Image Processing, 11, pp. 146-158, 2002. [14] M. K. Varanasi, and B. Aazhang, “Parametric generalized Gaussian density estimation,” The Journal of the Acoustical Society of America, 86, pp. 1404-1415, 1989.

IV - 666