In this chapter, a texture-based supervised segmentation algorithm derived la- belled objects from remotely sensed imagery. Texture was modelled with the joint distribution of LBP and local variance. The segmentation algorithm was a hierar- chical splitting technique, based on reducing uncertainty at the level of the image blocks that are obtained. By applying this technique, one does not only obtain a texture-based image segmentation, but also an indication of uncertainty for all object building blocks. The spatial distribution of uncertainty values provided information about the location and width of transition zones. This study showed that object uncertainty values provide important information to identify transition zones between fuzzy objects.
The proposed algorithm provided good segmentation results for a test case study with a composite image of five different textures. An overall accuracy of 96.20% was obtained. To illustrate the algorithm for mapping coastal objects, a LiDAR DSM and CASI image of a coastal area on the northwest coast of England were used. Good segmentation results were obtained for the extraction of land form objects from the LiDAR DSM, depicted by an overall accuracy of 86%. Un- certainty values provided meaningful information about transition zones between the different land forms. Land cover objects derived from the CASI image showed high uncertainty values and many incorrectly labelled objects. The overall accu- racy was 71%. The woodland area showed a characteristic texture in both data sources, however, the woodland object showed a different spatial extent and area in both segmentation results. This difference was caused by the occurrence of small patches of willow trees in, and on the border of, the woodland area. The texture of these willow trees is different from the pine trees in the area in the LiDAR DSM. The segmentation result of the LiDAR DSM correctly depicted the spatial extent of the pine area. However, the texture difference did not occur in the CASI image, resulting in a different segmentation result.
This and other segmentation errors can possibly be prevented by taking into ac- count spectral information from more than one band. The combination of textural and spectral information from all 14 CASI bands could greatly improve segmenta- tion results. This combination could be useful for mapping other land cover types in the area, like grasses, herbaceous plants, mosses, and shrubs. Additionally, the resolution of the neighbourhood set of the LBP measure affects the segmentation result. In this study, a neighbourhood set of the nearest eight neighbouring pixels (P= 8,R= 1) was used. A multi-resolution approach with different combinations of P and R might improve texture description.
Multivariate Texture-based
Segmentation
∗
For those who have seen the Earth from space . . . the experience most certainly changes your perspective. The things that we share in our world are far more valuable than those which divide us. D. Williams
6.1
A multivariate texture model
The LBP texture measure, as described in chapter5, allows a texture description of a single band. Most remote sensing images, however, consist of multiple bands. Segmentation of land cover objects based on a single CASI band in chapter 5
showed that unsatisfactory results were obtained. Including multiple bands might
∗This chapter is based on the following papers:
Lucieer, A., Stein, A. and Fisher, P. F. (in review ). Multivariate texture-based segmentation of remotely sensed imagery for extraction of objects and their uncertainty,International Journal of Remote Sensing. in review.
Lucieer, A., Orkhonselenge, T. and Stein, A. (2004). Texture based segmentation for identifi- cation of geological units in remotely sensed imagery,inA. Frank and E. Grum (eds),Proceedings of the 3rd International Symposium on Spatial Data Quality ISSDQ’04, Technical University of
6.1. A multivariate texture model
improve segmentation considerably, as a combination of bands provides more spec- tral information for identification of different land cover types.
In their psychophysical study Poirson and Wandell (1996) showed that colour and pattern information are processed separately by the human visual system. Mojsilovic et al. (2000) extracted colour-based information from the luminance and chrominance colour components. The achromatic pattern component was utilised as texture pattern information. Another approach is that of Panjwani and Healey (1995) which captured spatial relations both within and between colour bands with Markov random fields (MRFs). More recently, Pietik¨ainen et al. (2002) showed that the powerful LBP texture measure can also be applied to colour images. They processed colour information and texture information separately and obtained good classification results for colour texture images. Most research on colour texture is focused on images of different materials with a well-defined texture. In standard RGB-images the pattern in the different colour bands is often highly correlated. This makes it possible to summarise pattern information in a single band and process it separately from colour information. In remote sensing images information is recorded from different parts of the spectrum, therefore, texture in these bands is not necessarily similar. In between band relations should be taken into account when looking at multivariate texture measures for remotely sensed imagery. The LBP texture measure, described in chapter 5 is a robust, rotation invariant and flexible texture measure. An extension to the multivariate case is expected to provide good segmentation results.
In this chapter the new Multivariate Local Binary Pattern operator,M LBPc, is introduced and implemented. It is based on the univariate LBPc,j measure, describing local pixel relations in three bands, also known as colour texture. In addition to spatial relations of pixels within one band, pixel relations between bands are also considered. Thus, the neighbourhood set for a pixel consists of the local neighbours in all three bands. The local threshold is taken from these bands, which makes up a total of nine different combinations (figure6.1). This results in the following operator for a local colour texture description
M LBPc=P P−1
i=0 sign(gbi1−gbc1) +sign(gib2−gbc1) +sign(gib3−gbc1)+
sign(gb1
i −gbc2) +sign(gbi2−gcb2) +sign(gib3−gbc2)+
sign(gb1
i −gbc3) +sign(gbi2−gcb3) +sign(gib3−gbc3) (6.1) whereb1 is the first band,b2 is the second band, andb3 is the third band. The first part of the equation calculates LBP values for the center pixel of the first
R
G
B
gc,2
Figure 6.1: The neighbourhood set for the multivariate (three band) case not only takes into account the spatial relations within each of the bands, but also the relations between the bands.
band based on relations with the neighbors in the first band and the two other bands. The second part of the equation calculates LBP values for the center pixel of the second band and the third part of equation6.1 calculates LBP values for the center pixel of the third band. Each of the three central pixels is, therefore, compared with neighborhood pixels in the other bands. M LBPc is not just a summation ofLBPc,j of individual bands, it also models pixel relations between bands. These cross-relations can be important in the distinction of different color textures. A total of nine LBP values is obtained and summed to deriveM LBPc. The color texture measure is the histogram ofM LBPc occurrence, computed over an image or a region of an image. This single distribution contains 32P bins (for P = 8 resulting in 72 bins).
M LBPc measures the binary colour pattern of a texture. To complete this measure with contrast and variance information the colour histogram RGB-3D is