| GISdevelopment.net ---> Technology---> Image Processing |
Analysis of Window Size and Classification Accuracy using Spectral and Textural Information from JERS-1 SAR Satellite Image
Wikantika, K., M.Y. Andi Baso, Hadi F.
Expertise Group on Remote Sensing and Geographical Information Science (ReSGIS)
Department of Geodetic Engineering
Institute of Technology Bandung
Jl. Ganesha 10, Bandung 40132, INDONESIA
ABSTRACT
This study investigates the relationships between window size and classification accuracy using combination of spectral and textural features from JERS-1 SAR satellite image. The study focuses on extraction of entropy features using several window sizes which is calculated based on the gray level co-occurancy matrix (GLCM). A maximum likelihood supervised classifier is used for classification. The result shows that increasing the window size showed no significant contribution in improving the classification accuracy. In addition, results also indicate that the window size of 11x11 pixel is the optimal window size for classification, where the overall accuracy is equal to 80.00 % , the kappa index is equal to 0,780 as well.
INTRODUCTION
Spectral and textural features are a few of the basic elements of an image. Observed casually, spectral and textural features can look so different, but both are observable because of the nature and characteristic of an object recorded in an image. In remote sensing, spectral features are specific to every object that absorbs and reflects energy it receives, while texture characterizes the softness of an image surface. Important information of a texture can be obtained by extracting it from the image. Algorithmically, texture information of an image can be represented by a vector on every pixel. In this study, this feature will be extracted through a co-occurancy aproach represented by a gray level co-occurancy matrix (GLCM), also known as gray-tone spatial dependence matrix. [Harallick, et al., 1973].
Basically, the GLCM expresses the spatial relationship between a gray-level in a pixel with the gray-level in the neighboring pixels. The method used in this study is the entrophy method, which is one of the methods used to extract textural information using GLCM.
This study combines spectral features with texture extracted from different window sizes, from which the relationship between window size and classification accuracy will be analyzed, and from this analysis will be inferred the window size that gives the best accuracy.
The GLCM An image can be assumed to be rectangular in shape and have a number of Nx pixels in the horizontal direction and Ny pixels in the vertical direction. If the gray-level variation is to the level of Ng, then consider Lx = {1,2, …, Nx} as horizontal spatial area, Lz = {1,2, …, Ny} as vertical spatial area, and G = {1,2, …, Ng} as a set of gray level variation. Ly × Lx is the spatial resolution of an image. Image I can then be represented as a function that gives a few gray levels in G for every pixel or a pair of coordinates in Ly × Lx, I: Ly × Lx ? G [Harallick, et al. 1973], as indicated by Figure 1.
Figure 1.Gray-levels of a 4×4 image
The basic component of a conceptual frame about texture is size [Wikantika, et al. 2001]. Size is a matrix which elements denote spatial relations and gray level values of neighboring pixels. From this it can be said that this matrix is a function of angles and distance, where each gray-level value in a center pixel will have a spatial relationship with neighboring pixels in four directions: horizontal (0o), one vertical direction (90o) and two diagonal directions (45o and 135o). To depict this array, we must invoke the limitations of the array, including the outer area of the image. Harallick et.al 1973) considers that textural information in an image I is contained in all or parts of the spatial relationship that the gray level values have. More specific, they assumed that the textural information is determined only by the relative frequency matrix Pij with two closest pixels, separated by a distance d on the image, one with the gray level value of i and the other one with gray level value of j.
Figure 2. Angle and distance function of a GLCM
From Figure 2, we know that the observed pixel is a pixel with BV = bv5, which has spatial relationship with bv4 and bv6 in direction 0o, bv3 and bv7 in direction 45o, bv2 and bv8 in direction 90o, and bv1 and bv9 in direction 135o.
This reseach is focused more on the use of the entropy method to extract and classify the important information of the texture. This is based on the method's ability to classify an object with higher level of accuracy compared to other methods [Wikantika et.al, 2004]. Literally, entropy is the chaotic degree, which means that pixels with high entropy will have high degree of chaos [Baraldi and Parmiggiani, 1993; Haralick et al., 1973].
In this entropy method, we use the following formula [Baraldi and Parmiggiani, 1993; Haralick et al., 1973]:
Geometric and radiometric correction
Geometric is carried out to eliminate geometric distortion (the non-systematic properties) to obtain relations between image coordinate system (row, column) with projection coordinate system. The process of geometric correction of this JERS-1 SAR image refers to a rectified Landsat ETM image, not to a digital topographic map. This is because the difficulties of identifying specific details from JERS-1 SAR image. Radiometric correction consists of filtering to reduce speckles from the image. Filtering is done by using Lee Method with window size 5×5.
Texture extraction using GLCM
Texture extraction is done through GLCM-based Entropy Method, which justified by the method's ability to classify an object with higher accuracy than other methods [Wikantika, et.al, 2004]. From the rectified JERS-1 SAR image, seven images of texture extraction result will be produced as the window size has been divided in to window size 3×3, 5×5, 7×7, 9×9, 11×11, 13×13, and 15×15 (Figure 3)
Figure 3. Texture extraction process in 7 window sizes
Texture and spectral combination
After the stage of texture extraction is finished and seven texture images has been produced, the next step is combining spectral image with texture extraction image, as shown in Figure 4.
Figure 4. Spectral and textural combination
Image classification and accuracy assessment
This study uses supervised classification with maximum likelihood method [Lillesand and Kiefer, 2000]. The flow for this classification is shown in Figure 5.
Figure 5. Classification using maximum likelihood method
Because of the difficulties in identifying specific details in the combined images, the training site assignments were done by taking the training results from images who have spectral information only (Figure 5). This can only be done if both of the images are guaranteed to have the same datum dan coordinate system.
As for the accuracy assessment of classification result of combined spectral features and seven images from texture extraction, we used reference data from year 2003 Quickbird image. Even tough the reference data, taken in the year 2003, and the JERS-1 SAR image, which taken in the year 1997, have a significant difference in record year, this can be anticipated by using a sample point only in the areas that are assumed not have significant changes. This can be seen in Figure 6.
Figure 6. Accuracy assessment of classification results
From the combined images with different sizes, sample points were taken by stratified random sampling. From the four existing classes, we took 10 sample points from each class, so we have a total of 40 sample point and every class will have the same chance to be taken in this accuracy assessment. The advantage of the stratified random sampling method is that all kind of land will be included in the sample, no matter small it is.
From the accuracy assessment we obtain results as follows:
- For window size 3×3, overall accuracy is 37.50% and the kappa index is 0.338.
- For window size 5×5, overall accuracy is 47.50% and the kappa index is 0.441.
- For window size 7×7, overall accuracy is 62.50% and the kappa index is 0.596.
- For window size 9×9, overall accuracy is 77.50% and the kappa index is 0.753.
- For window size 11×11, overall accuracy is 80.00% and the kappa index is 0.780.
- For window size 13×13, overall accuracy is 75.00% and the kappa index is 0.723.
- For window size 15×15, overall accuracy is 62.50% and the kappa index is 0.588
Figure 7. Relations between window sizes with kappa index
Figure 8. Relations between window sizes with overall accuracy
From Figure 7 and 8, the kappa index and overall accuracy in window size of 3×3 through 11×11 is increase, but decrease in window size of 13×13 and 15×15. From this fact we can inferred that increasing the window size in the extraction process can only increases kappa index and overall accuracy to a certain limit, that is on window size 11×11. It is also can be concluded that increasing the window size does not provide significant contribution to increase the accuracy of the classification using combination of spectral and textural features. From this accuracy assessment, we also obtained that the best window size for image classification is 11×11 with 80.00% overall accuracy and kappa index 0.780.
CONCLUSIONS
From this study it can be concluded that:
- Increasing the window size during the texture extraction stage does not give significant contribution in improving the classification accuracy.
- From the seven window size that has been evaluated, the best window size for classification is 11×11 with 80.00% overall accuracy and kappa index 0.780. Window size 7×7, 9×9, 13×13, and 15×15, are still usable from classification since the overall accuracy is above 50.00%
References
- Baraldi, A., and Parmiggiani, F. (1995). An Investigation of the Textural Characteristics Associated with Gray Level Co-occurrence Matrix Statistical Parameters. IEEE Transactions on Geoscince and Remote Sensing volume 33
- Haralick, R. M., Shanmugam, K., and Dinstein, I. (1973), Textural Features for Image Classification.IEEE Trans. Sys. Manage. Cybernetics, SMC-3(6), 610-621
- Lillesand, T. M. and Kiefer, R. W. (2000). Remote Sensing and Image Interpretation. (4th Edition). Jhon Wiley and Sons, New York
- Wikantika, K., Uchida, S., and Yamamoto, Y., (2001), Discrimination of Vegetable Field in Mountainous Area with Spectral and Textural Information Derived from Landsat-ETM, Proceedings of the International Symposium on Land Use-Land Cover Changes Contribution to Asia Environmental Problems, Tokyo, Japan
- Wikantika, K., Uchida, S., & Yamamoto, Y. (2004). An Evaluation of The Use of Integrated Spectral and Texture Features to Identify Agricultural Land Cover Types in Pangalengan, West Java, Indoensia. Japan Agricultural Research Quarterly, Vol. 38, No. 2., 137-148
| Window 3x3 | Window 5x5 |
 |
 |
| Window 7x7 | Window 9x9 |
 |
 |
| Window 11x11 | Window 13x13 |
 |
 |
| Window 15x15 | |
 |