Texture Feature Extraction for Land-cover Classification of Remote Sensing Data in Land Consolidation District Using Semi-variogram

Transcription

1 Texture Feature Extraction for Land-cover Classification of Remote Sensing Data in Land Consolidation District Using Semi-variogram Anzhi Yue, Su Wei, Daoliang Li, Chao Zhang *, Yan Huang College of Information & Electronics Engineering, China Agricultural University, Beijing , China. *Corresponding author: Chao Zhang (Vice-Prof.) Tel: , Abstract: The area of land consolidation projects are generally small, so remote sensing images used in land cover classification are generally of high resolution. The spectral characteristics of the high-resolution remote sensing data are unstable, while the texture feature is prominent. In view of this issue, this paper study the spatial relation between the adjacent pixels in the remote sensing image, and selected the lag distance of the semi-variogram that is determined when the value of the semi-variogram tended to be constant as the cooccurrence window size. Sometimes the window size is the most important influencing factor in the texture feature extraction process. Moreover, under the restraint of the classification results, this paper introduces a method to compute the co-occurrence features with a timely changeable co-occurrence window size according to the semi-variogram analysis. This paper takes Zhaoquanying land consolidation project located at Beijing Shunyi District as an example, the texture feature is extracted from SPOT5 remote sensing data of land consolidation project area in the TitanImage secondary development environment. The results show that the classification has improved. Key-Words: Semi-variogram, Land Consolidation, Texture feature, Classification 1 Introduction Classification plays an important role in the information extraction for land consolidation project. The general classification approaches (including supervised classification and nonsupervised classification) are no longer suitable for the high-resolution remote sensing data (SPOT5 2.5m, IKNOS 1.0m, QuickBird 0.61m) due to spectral instability of high-resolution images. The general approaches greatly depended on the spectral information, and neglected some important information in the high-resolution image, such as texture feature, shape features, et al. Therefore, these classification approaches cannot be used in the high spatial resolution images. A new classification approach should not only depend upon its spectral information, but also uses other important information in the image. Joining the texture feature to normal classification approaches have already become a hot research issue in the domain of image information extraction. Co-occurrence matrix is the most common and widely method used in the statistical texture analysis, through which the spatial relationship of the pixels is researched to describe the remote sensing image. In order to improve classification, a new method of classification combining texture feature based on semi-variogram is proposed in this paper. 2 Related Works and the Proposed Method 2.1 Co-occurrence Matrix A co-occurrence matrix is a square matrix whose elements correspond to the relative frequency of occurrence of pairs of gray level values of pixels separated by a certain distance in a given direction [1] [2]. As shown in Fig.1, where ( Δx, Δy ) represents distance between two pixels, i and j are the DN value of pixels, d is the distance between two pixels. The gray level co-occurrence matrix is defined as: P( i, j, d, θ )( i, j = 0,1,2,, N 1) (1) Where the number of gray series is N, θ is the angle between the lines defined by two pixels and X-axis. ISBN: ISSN:

2 i d x y X j Y The gray level co-occurrence matrix contains comprehensive information of image distribution, such as direction, partial neighborhood, changing range, but which cannot be used to extract region texture directly. Therefore, in order to describe texture feature using matrix, extracting texture feature from matrix is necessary. Many methods, such as homogeneity, contrast, mean, angle secondorder moment, have been proposed to measure textural properties for statistical textural analysis using the gray level co-occurrence matrix [3]. 2.2 Semi-variogram The semi-variogram function describing the spatial variation of samples as a function their apart distance is an intermediary in geostatistics calculation [4]. The semi-variogram was firstly put forward by G.Matheron in 1960s [5]. The theoretical foundation and study object of semi-variogram are region variation distributed rule in spatial range and time-spatial range [6].To describe the semivariogram spatial variant structure of the remote sensing data, the mathematic structure model of the semi-variogram is defined as: N ( h) 1 2 ( h) [ Z( xi ) Z( xi h)] 2N( h) i 1 (2) Where (h) represents the values of semivariogram, while h is the separation distance Fig.1 Gray level co-occurrence matrix between two samples, N (h) is equal to the number of value pairs in which the separation distance is equal to h Z( x, i ) x is calculated using i, which is corresponding to DN value in the remote sensing image. Semi-variogram function is typically expressed in semi-variogram curve. The semi-variogram curve indicates the relationship between semi-variogram function values (h) and sample distance h. Nugget, base value and range are three important parameters in semi-variogram s theoretical model. 2.3 Semi-variogram-based Texture Feature Windows Determined When extracting texture feature by co-occurrence matrix, main parameters used includes texture feature, gray series, direction and separation distance between two pixels, window size. F. Dell Acqua et al. extracted texture feature by using cooccurrence matrix and semi-variogram analysis for mapping urban density classes in satellite SAR data. The results show that the joint use of co-occurrence textural features and semi-variogram analysis to optimize co-occurrence window size can be, in terms of, as effective as a long exhaustive search of the best scale [7]. Textural analysis and results for fixed window size is given in Fig. 2. (a) 3*3 (b) 7*7 (c) 11*11 Fig.2 Heterogeneity extracted using fixed textural window As shown in the Fig. 2, the textural window size affected the textural features extraction to a certain extent [8]. When textural window is relatively small, textural information extracted is fragmentized, more details are manifest and more spots exist in textural information; on the other hand, the information of small target may be filtered and the boundary of object is fuzzed, so that also can not acquire the ISBN: ISSN:

3 satisfied results. When the textural features are extracted by using fixed window in classification, the overall classification can improve to a certain extent, however, with the window gradually increasing, the overall classification begin to reduce. Window size for classification of the different categories is not same. When window size smaller than 9*9 pixels is used, the classification of the farmland is gradually increasing, reversely, the classification of woodland is gradually decreasing. This paper studies the spatial relations between the adjacent pixels in the remote sensing image according to semivariogram value of samples, and selects the lag distance of the semi-variogram. The value of selected semi-variogram tended to be constant to the co-occurrence window size, which is the most important influencing factor in the textural features extraction process [9]. 3 Study Site and Data Preparation The study site is Zhaoquanying town which located in Shunyi district, Beijing where a land consolidation project exists. This area is characterized by rice-farmland and flourishing vegetation. The data used in study is SPOT5 remote sensing images, The SPOT5 image consists of a panchromatic band with 2.5m spatial resolution and four multispectral bands(i.e. near-infrared(nir), red, green, and Short Waved-length Infrared) with 10m spatial resolution, which were acquired in 2006 without any clouds/hazes. All bands of imagery were geo-referenced to a Transverse Mercator projection and Krasovsky spheroid with an RMSE of 1 pixel. The SPOT5 remote sensing image had to be preprocessed, including geometric calibration to the multi-spectral band data with the panchromatic band data and HSV fusion for the results after correction with panchromatic data, and the resolution of the final result is the same as that of the panchromatic band data. 4 Texture Feature Extraction Process The flow graph of texture feature extraction is given in Fig.3. Select a same region with the rich textural information as a research object from the fused data and the panchromatic band data. Furthermore, cut out the training sets as a sample (ROI) for the different types of the land-use, then convert it to ASCII format in ENVI4.3, join the line-column ranks corresponding to DN value in Office Excel and saved as.txt format. The.txt data will provide information for semi-variogram analysis finally. Meanwhile, the fused image was classified with the supervised classification method of Maximum Likelihood and the result was corresponded with the result of the window size analysis as constraint of the panchromatic band data to compute the cooccurrence features. Original image Supervised classification Edit ROIs Classification image ROI files Post classification Computer semi-variogram value, determine window size Constraint condition Extract texture feature Figure of texture feature 4.1 Semi-variogram Analysis Fig.3 texture feature extraction ISBN: ISSN:

4 All class samples will be carried out semi-variogram analysis respectively in GS + software in order to obtain the window size of co-occurrence matrix suit to the specific class and determine the size of textural window. However, uncertainty exists in selection for all the classes in land-cover classifications, so the size of the ground object must be considered comprehensively when the size of samples is determined [10]. When sample range is relatively big, the interference of none-sample object will be strengthened; in contrast, sample object will be destroyed, both cannot acquire precise semivariogram features of land-use classes. This paper, farmland s semi-variogram function curve is given in Fig. 4, and textural analysis window to every class determined finally is shown in Table 1. Semi-variogram analysis 10 Semi-variance Sample 1 Sample 2 Sample 3 Sample 4 Sample 5 Sample 6 Sample Separation Distance Fig.4 semi-variogram function curve of farmland Table 1 HOM for different textural window classes Size of textural window farmland 1 9*9 farmland 2 9*9 woodland 5*5 water 3*3 construction 11* Texture Feature Extraction Base on the size of textural window for each class in the land consolidation, the fused image is classified with the supervised classification method of Maximum Likelihood. The result is corresponded with the result of the window size analysis of the above, which is set as constraint of the panchromatic band data to compute the co-occurrence features. When extracting texture feature, the initial classification type of image is recorded timely in order to change the window size of every pixel in the extraction process. Result of non-fixed window for texture feature extraction is given in Fig. 5. As shown in Fig.5, texture feature extracted using non-fixed window reduces the Pepper-Salt phenomenon, and guarantees the details of image and clarity of edge compared with results extracted using fixed window given in Fig.2. Fig.5 Mean features extracted using non-fixed textural window 5 Classifications and Result Discussion Texture feature image extracted will be combined with the band fused images as a band, which will ISBN: ISSN:

5 acquire the Maximum Likelihood supervised classification image in the same training area. Meanwhile, the original image and the fixed window image will be classified with the Maximum Likelihood supervised classification method. The classification result is shown in Fig.6. (a) original image (b) 3*3 (c) 7*7 (d) 11*11 (e) non-fixed window Fig.6 classification images involved to texture mean features of each textural window As shown in the Fig.6, there are a lot of farmland classified into woodland. This error is due to the similarity of farmland and woodland in the spectra features [12], but the texture feature between farmland and woodland is different, this problem is obviously improved by using the texture feature, which was extracted from the 3*3 fixed window. However, Pepper-Salt phenomenon is serious problem due to small textural window, especially in some construction sites. With the increasing of textural window, Pepper-Salt phenomenon has improved, but filtered out some details at the same evaluation features time [13]. In this paper, the method of extracting texture feature using non-fixed window is adopted. It not only assures the information of farmland not be wrong classified into other classes, but also can meet the need of visual investigation. Finally, the overall and the KAPPA coefficient are estimated for every classified image. The overall for classification of fused original image is 80.96%, Kappa 0.68, and other evaluation results of classification image are shown in Table 2. Table 2 Accuracy analysis Angle second-order moment Homogeneity Mean Relativity Overall Kappa Overall Kappa Overall Kappa Overall Kappa methods Fixed window Non-fixed window The maximal, which is chosen from the classification of the 3*3 pixels window to the 11*11 pixel window involved in the classification, is set as results given in the Table.2. Taken the mean features for example, the 3*3 pixel window of the classification is 81.08%, 7*7 is 78.02%, and 11*11 is 77.57%. So the classification of the 3*3 pixel window was chosen as the final value of the mean feature of the fixed window. The experimental result demonstrated that the method of adding general textural feature to the classification can improve the classification to a certain extent. And the new extract method of textural feature in this paper can improve ISBN: ISSN:

6 the of the classification further. For instance, in regard to the textural measure of mean, the overall of the original fused image classification result was only 80.96%, while the overall of the classification image by adding the non-fixed window size texture feature raised to 86.02%, and the overall of the classification image by adding the fixed window size texture feature only achieved 82.80%; the KAPPA coefficient of the original fused image classification result is 0.68, the KAPPA coefficient of the classification image by adding the texture feature with non-fixed window size raises to 0.75, while the KAPPA coefficient of the classification image by adding the texture feature with fixed window size achieves Therefore, it is obvious that the of the new method developed in this study is sufficient for practical investigations. 6 Conclusion and Future Works This paper proposes a method by adding non-fixed window extract texture feature for remote sensing image classification, the results demonstrated that the method can make full use of textural information in high-resolution remote sensing image in the land consolidation, especially for the farmland with rich textural information which can play an important part in discriminating it from other green plants. Comparison of the results using the non-fixed window size showed that appropriate window size should be neither small nor large. Because small window size is possible to make the texture feature fragmentized, while large window size may filtered information of small target, and low efficiency problem will be produced by using large window size. The proposed method in this paper can be effective and sufficient for high-resolution classification. Even though the idea of the method is simple, the stable textural window size is difficult to determine for each class due to complexity and diversity of the same land feature. It required a large numbers of samplings and semi-variogram analysis to improve the classification. Therefore, the actual operation may be relatively complex. In conclusion, the method proposed is still in the research stage and need further improvement. Further work will consist of the combination of the proposed method with the pixel-based classifier and object-oriented classifier, which integrated different textural methods that will be more effective for the land consolidation project. Acknowledgement This work has been supported by The National High Technology Research and Development Program of China (contract number: 2006AA12Z129). References: [1] Haralick R M, Textural features for image classification, IEEE Trans SMC, Vol.3, No.6, 1973, pp [2] Chuen-Lin Tien, You-Ru Lyu and Shiao-Shan Jyu, Surface flatness of optical thin films evaluated by gray level co-occurrence matrix and entropy, Applied Surface Science, In Press, Corrected Proof, Available online 26 January 2008 [3] T.Randen,J.H.Husoy, Filtering for Texture Classification: A Comparative Study, IEEE Trans. PAMI, Vol.21, No.4, 1999, pp [4] Peter I. Brooker, Changes in dispersion variance consequent upon inaccurately modelled semivariograms, Mathematics and Computers in Simulation, Vol.30, No.1-2, 1988, pp [5] Matheron G. M, Principles of geostatistics, Econ. Geol., No.58, 1963, pp [6] Eulogio Pardo-Igúzquiza, Peter A. Dowd, VARIOG2D: a computer program for estimating the semi-variogram and its uncertainty, Computers & Geosciences, Vol.27, No.5, 2001, pp [7] F. Dell Acqua, P. Gamba, G. Trianni,Semiautomatic choice of scale-dependent features for satellite SAR image classification, Pattern Recognition Letters, Vol.27, 2006, pp [8] Timothy C. Haas, Kriging and automated variogram modeling within a moving window, Atmospheric Environment, Vol.24, No.7, 1990, pp [9] D. Puig, M.A. García, Determining Optimal Window Size for Texture Feature Extraction Methods, IX Spanish Symposium on Pattern Recognition and Image Analysis, Vol.2, 2001, pp [10] Ferro, C.J, Warner, T.A, Scale and texture in digital image classification, Photogrammetric Engineering and Remote Sensing, Vol.68, No.1, 2002, pp [11] M. Park, J.S. Jesse, L.S. Wilson, Hierarchical Indexing Images Using Weighted Low Dimensional Texture Features, IAPR Vision Interface, Calgary, 2002, pp [12] J.G. Zhang, T.N. Tan, Brief review of invariant texture analysis methods, Pattern Recognition, Vol. 35, No.3, 2002, pp [13] M. Pietikainen, T. Ojala, Z. Xu, Rotation- Invariant texture classification using feature distributions, Pattern Recognition, Vol. 33, 2000, pp ISBN: ISSN: