2008 International Conference on Advanced Computer Theory and Engineering Region Growing Tree Delineation In Urban Settlements LAU BEE THENG, CHOO AI LING School of Computing and Design Swinburne University of Technology Sarawa Campus Kuching, Sarawa, MALAYSIA E-mail: blau@swinburne.edu.my Abstract Region growing has been widely used in image segmentation applications. Iovan et al. (2007) has used region growing technique to delineate trees in urban settlements with Quicbird imagery. However, there is a room for improvement in the extraction accuracy. This study attempts to improve the tree crown detections from high resolution satellite imagery with wavelet transform and region growing. We shall discuss how the biorthogonal wavelet transform is utilized in the segmentation first, and then follow by the grass/tree segmentation by region growing. It is proven that accuracy is higher when wavelets are integrated into the segmentation and extraction procedures. 1. Bacground Generally, satellite imagery provided by Quicbird provides higher spatial resolution compared to other provider lie Ionos. Higher spatial resolution data in terms of sizes, high frequency spatial arrangements, high class spectral variances and similarities of obects in satellite imagery have demanded better automated spatial structure processing. However the improvement of the spatial resolution in satellite imagery lie Quicbird does not automatically transform into a better classification accuracy of multi-spectral classification. The spatial structures retrieval problem remains. Spatial structures lie trees, buildings and roads need to be adequately digitized, mapped or integrated into practical systems lie Geographic Information System with improved algorithms as well. Satellite imageries for urban settlements are composed of diverse structures. For instance, buildings, roads, open ground, fields and trees. With the finer spatial resolution provided by Quicbird, the spatial obects have become more visible. Though the spatial obects are more distinguishable by human eyes visually, it complicates the spectral separability in spatial obect detection system. For instance, weeds, grasses and trees may have same colour frequency. On the other hand, looing at the bright side of the resolution improvement, spatial structures semantic information actually forms a hierarchy of scales. Structural texture is an outcome of obects being arranged in such hierarchy. Hence, incorporation of multi spectral textures can improve the detection of spatial structure/obect lie tree crowns, buildings and roads from Quicbird satellite imagery. Selection of texture in sub-bands can be done based on the mean, entropy, variance and angular moment. Hence the selected sub-bands are distinguishable between tree crowns and similarly reflecting features such as grass and open ground. In short, texture analysis is the direct embodiment of the obect s structure and spatial arrangement in satellite imagery. This implies that texture is a ey factor of spatial resolution. In Iovan et al. (2007), a region growing method is used to delineate tree crowns from other spatial structures in satellite imagery. However, the delineation rate (only 78% out of 41 tree crowns from ground truth) can be further improved. In order to improve detection rate of the tree crowns, multi-scale wavelets decomposition is proposed for extracting the vertical, horizontal and diagonal texture components. Then the wavelets are extracted by region growing method. 2. Proposed segmentation model For a given region of interest that is a vegetation area contains greens lie trees and grasses, a texture analysis is conducted to differentiate grasses from trees. In wavelet transformed imagery, treed areas are characterized by higher wavelet variance compared to grasses. The method developed to separate grass from trees taes into account this property by computing the local variance on the wavelets. The resulting image is threshold to obtain mass for grass and treed areas. Consequently, a region growing method is use to segment the imagery by wavelets. 978-0-7695-3489-3/08 $25.00 2008 IEEE DOI 10.1109/ICACTE.2008.140 915
2.1 Structural approach Structural approaches as compared to statistical approaches have a significant benefit that is deterministic textures are measured. They find a basic region of a texture for classification and measure the perceptual similarity through spatial frequency-based wavelet transforms. This is because wavelet transforms of an image captures localized spatial and frequency contents. As image texture is a function of scale, another benefit of wavelet decomposition is the unified framewor for multi-scale texture analysis. Wavelets are better over Gabor transform and Fourier analysis because wavelets present features at the most probable representation scales by varying the spatial resolution. Furthermore wavelets can be changed to suit various applications. 2.1.1 Biorthogonal wavelets A commonly used signal processing strategy is called multi-resolution analysis, MRA (Zhang et al. 1991). It comprises of a set of specialized filters designed to extract features occurring frequencies and temporal localization in different resolutions from images. For satellite imagery, the frequency information can be identified by colour variation whereby the temporal localization can be identified through spatial localization of the colour variation. In this matter, a wavelet transform function measures the grey-level colour variations at different scales. Contours of structures lie roads, buildings or tree crowns correspond with significant colour contrasts and can be distinguished from the local maxima of a wavelet transform. This is done through simultaneous time-scale representation of the signal with biorthogonal wavelet (BW). With BW, wavelets with multi spectral that are not possible with orthogonal wavelets can be included for analysis. BW transform for the structure variations are obtained by a set of dilated and translated scaling functions nown as the mother wavelet (Mathur 2002). The mother wavelet function is in Equation 1. / 2, ( n) 2 (2 n ) (1) These basic function at level () is expressible as a linear combination of the wavelet function at the next level (+1) in Equation 2 where g(m) is the wavelet filter., ( n) g( m 2) 1, m ( n) m (2) In a continuous system, Equation 2 can be represented as its wavelet derivatives (Burrus et al. 1998) as in Equation 3. f ( n) c 0 ( ) ( n) d ( ) ( n (3) 0, ) 0, The fast discrete Wavelet Transform (fdwt) can be expressed as in Equation 4 (Hong and Zhang 2004) in which the wavelet filter [g*(n)] is now a high-pass filter extracting the detail coefficients d + 1 (). d 1 ( ) c ( n) g ( n 2) (4) n * All coefficients are down sampled with dyadic fdwt where only every other coefficient is taen. The implementation of the fdwt are for level (+1) approximation, horizontal, vertical and diagonal subimages. They correspond to four wavelet coefficient channels comprise of the low low, low high, high low and high high. 2.2 Tree crown detection For tree crown detection, a region growing method is used. The prerequisite for region growing method is seed point identification for each region before growing. Performance of the region growing algorithm is highly dependent on the number of seed points initializing each region. Ideally, one seed point per region is the best. 2.2.1 Seed point initialization For each tree crown, having one seed point for each tree is desired. To reduce the number of possible candidates for a tree crown, Gaussian filter is used as smoothing filter with a predefined template. Local maxima for Gaussian on this image represent the seeds for the region growing algorithm. The template has the average size of the trees in the satellite imagery. To determine tree crowns, maximum diameters of the tree crowns and all points having the same width are treated 916
as tree crowns. In the first iteration, points corresponding to the tree crowns with maximum diameters. Iteratively the analysis altitude, w is decreased. At each step, all points at higher widths than w are surveyed. A new seed is piced up when a new region appears. This does not affect pixels previously detected and initialized as seeds. 2.2.2 Single seed region growing Each of the previously detected seeds is grown to become a region. A priority queue is established for the order the seeds are processed: seeds which are border pixels to a region are processed sooner than seeds which are not border pixels. The higher the value the border pixel is the sooner it will be connected to a region. This value is taen from a new image, a random wal image which is obtained from the original image by simulating random wals for each seed point. The value of each pixel represents the number of times the simulated particles have reached the pixel. A series of constraints decide on the rapidity of a pixel aggregation to a region. Figure 1. Burlington s partial urban settlements contain buildings, roads, walways, bushes, open ground and water (QuicBird 2007) 2.3 Data For the most informative band for urban-tree texture extraction in this study, different tree species of various spectral characteristics representing the urban trees and non-tree areas lie buildings, roads, walways, bushes, open ground and water were selected as training sets for QuicBird bands: blue, green, red and Near Infrared at 2.44 m resolution, and panchromatic at 0.61 m spatial resolution. Figure 1 shows the original resolution of Burlington urban settlements and Figure 2 shows the raster map being transformed to a color ramp based on wavelets of RGB bands. Table 1. Panchromatic, spectral and pansharpened bands and spatial resolutions of Quicbird imagery Band Wavelengt h (mm) Original resolution (m) Blue 0.45 0.52 2.44 0.61 Green 0.52 0.6 2.44 0.61 Red 0.63 0.69 2.44 0.61 Near infrared 0.76 0.9 2.44 0.61 Panchromatic 0.45 0.9 0.61 - Pansharpened resolution (m) Figure 2. The transformed Burlington s partial urban settlements 3. Evaluation The approach used for the evaluation taes Iovan et al. (2007) as the bench mar. It was a purely region growing tree crown delineation. Hence it was used as a reference to determine whether wavelets improve the delineation. Statistical analysis was conducted for analyzing total number of trees in the ground truth, omitted trees and under segmented trees. The spatial analysis of the segmentation included complete segments, over segmented and under segmented trees. Complete segments were the correctly identified trees. A complete segment corresponded to one and only one 917
segment in the ground truth and vice versa, with a matched area greater than 80%. Over segmented tree crowns were those cases where more than one segment was associated with the ground truth. Under segmented tree crowns were segments which had significant part that was greater than 10% of a tree. 4. Findings The proposed wavelet region growing for detecting tree crowns from all other greens produce better results as shown in Table 2. This is to compare with the benchmar by Iovan et al. (2007) which uses region growing alone for detecting vegetation and nonvegetation areas to delineate the trees from greens. Table 2. Comparison between the results obtained for tree crown delineation Figure 3. Delineated trees from urban settlements Benchmar Proposed Measure by tree Qty. % Qty. % crowns Correctly segmented 193 78.78 238 97.14 Over segmented 8 3.26 0 0 Under segmented 22 8.98 0 0 Omitted 22 8.98 7 2.86 Total detected 223 91.02 238 97.14 245 245 For obtaining tree crowns ground truth data, manual delineation was done. In comparison with the ground truth data, 97.14% of the tree crowns surface in the reference delineation was correctly classified as trees with the automated region growing segmentation with wavelet transformed satellite imagery. The main reason causing omitted tree is the tree crown diameter that is unusual i.e. lower to average. The number of correctly segmented trees is higher for the proposed method due to the way the wavelets transformation. The results of the proposed method are improved by 18.36% when it uses wavelet transformation for region growing segmentation. These results show the potential of wavelets in delineating trees from urban settlement with higher accuracy. Figure 4. Delineated trees from urban settlements shown in actual Quicbird imagery 5. Conclusion This research has conducted improvements on region growing that is widely used in image segmentation applications. The benchmar is based on Iovan et al. (2007) to delineate trees in urban settlements with Quicbird imagery. This study improved the tree crown detections from high resolution satellite imagery with wavelet transform and region growing. Biorthogonal wavelet transform was utilized in the segmentation. The research has proven that accuracy is higher when wavelets are integrated 918
into the segmentation and extraction procedures of tree crowns in urban settlements. 6. References [1] Burrus, C. S., Gopinath, R. A. and Guo, H. (1998). Introduction to Wavelets and Wavelet Transforms. New Jersey: Prentice-Hall. [2] Digital Globe Inc. QuicBird Sample Imagery. Accessed 30/03/06. http://www.digitalglobe.com/sample_imagery.sht ml. [3] Gruen, A. (2000). Potential and Limitation of High-Resolution Satellite Imagery. Proceedings of the Asian Conference on Remote Sensing, December 4-8, Taiwan. [4] Hong, G. and Zhang, Y. (2004). The effects of different types of wavelets on image fusion. Proceedings of the International Society of Photogrammetry and Remote Sensing, Istanbul, pp. 915 920. [5] Iovan, C., Boldo, D. Cord, M. and Erison, M. (2007). Automatic extraction and classification of vegetation areas from high resolution images in urban areas. In B. K. Ersboll and K. S. Pedersen (Eds.): SCIA 2007. LNCS4522, pp. 858 867. Springer-Verlag. [6] Mathur, A. (2002) Dimensionality reduction of hyper-spectral signatures detection of invasive species. MSc Thesis, Mississippi State University. [7] Ouma, Yashon O., Ngigi, T. G. and Tateishi, R. (2006). On the optimization and selection of wavelet texture for feature extraction from highresolution satellite imagery with application towards urban-tree delineation. International Journal of Remote Sensing, 27:1, 73 104. [8] Thomas, N., Hendrix, C. and Cogalton, R.G. (2003). A Comparison of Urban Mapping Methods Using High-Resolution Digital Imagery. Photogrammetric Engineering and Remote Sensing, Vol. 69, No. 9, September 2003, pp. 963-972. [9] Toutin, T. and Cheng, P. (2002). 3D Models for High Resolution Images: Examples with QuicBird, IKONOS and EROS. Symposium on Geospatial Theory, Processing and Applications, Ottawa. [10] Zhang, W., Hasegawa, A., Itoh, K. and Ichioa, Y. (1991). Image processing of human corneal endothelium base on a learning networ. Applied Optics, Vol. 30, pp. 4211 4217. 919