04-130 3/14/06 9:05 PM Page 373 Three New Implementations of the Triangular Prism Method for Computing the Fractal Dimension of Remote Sensing Images Wanxiao Sun Abstract Based on Clarke’s (1986) triangular prism concept, this paper proposes three new methods to compute the fractal dimension (D) of remote sensing images. Our first method involves searching a pixel on each edge of a square whose digital numbers (DN) value has the largest deviation from the central pixel. Our second method uses a pixel on each edge of a square whose DN deviation from the central pixel is closest to the mean DN deviation from the central pixel to all pixels on the same edge. In our third method, eight pixels on the four edges of a square are used. Furthermore, common to the three proposed methods is the use of actual DN of the central pixel. The proposed computation methods have been tested using both simulated fractal surfaces and real images. Results show that the proposed methods appear to generally perform better than Clarke’s 1986 method for synthetic images with complex textures. Introduction Fractal-based texture analysis of remotely sensed images has generated considerable interest in the remote sensing community in the past two decades (Pentland, 1984; De Cola, 1989; Lam, 1990; De Jong and Burrough, 1995; Myint, 2003). A main reason for this increased interest in the fractal model seems the realization that incorporation of spatial information in digital image analysis can help improve classification accuracies (Haralick et al., 1973; Pratt et al., 1978; Gong and Howarth, 1990). As a relatively novel class of spatial technique, fractal analysis appears to provide considerable potential for characterizing textures in remotely sensed images. Fractal geometry was introduced and popularized by Mandelbrot (1977) to model natural shapes (e.g., coastlines and terrain) as well as other complex forms that traditional Euclidean geometry fails to analyze. A major application of fractal geometry in the geosciences has been the use of fractal dimension to characterize the form of environmental phenomena at different scales (Goodchild, 1980; Mark and Aronson, 1984; Goodchild and Mark, 1987). The fractal dimension, often denoted as D, is a central construct of fractal geometry. It is called fractal dimension because it is a fractional (or non-integer) number. A coastline’s fractal dimension, for example, can take on any non-integer value between 1 and 2, depending on the degree of irregularity of its form. The more contorted a coastline is, the higher its fractal dimension. Similarly, a terrain surface’s fractal dimension may be a non- integer value between 2 and 3. As such, the fractal dimension can be thought of as a parameter capable of capturing the geometrical complexity of an object being analyzed. A remotely sensed image can be viewed as a hilly terrain surface whose “elevation” is proportional to the image gray value. Technically, an image can be interpreted as a 3D space where the x, y coordinates represent 2D position on the image plane and the z coordinate represents the gray level values or digital numbers (DN). Most remotely sensed images are spatially and spectrally complex. As such, the fractal dimension appears to be a useful parameter for measuring the surface roughness (i.e., brightness differences) of remotely sensed images. Several studies have used fractal techniques to characterize textures and features in digital images. For example, Pentland (1984) has shown that computed D values are useful in edge detection, image segmentation, and other image analysis applications. De Cola (1989) used fractal techniques to examine the scaling characteristics of feature classes constructed from Landsat TM images. Lam (1990) demonstrated that different land-use types exhibit textures of different D values. De Jong and Burrough (1995) presented a “local D algorithm” to classify Mediterranean vegetation types in remotely sensed images. Myint (2003) has recently compared the accuracies of image classification using fractal methods as well as other texture analysis techniques. The utility of D as a texture measure depends to a large extent on the availability and accuracy of the methods of computing D. Several researchers (e.g., Tate, 1998; Lam et al., 2002) have emphasized that developing efficient and reliable algorithms is key to the application of fractal techniques to digital image analysis. The triangular prism method was proposed by Clarke (1986) to primarily compute the fractal dimension of topographic surfaces. Despite this, Clarke’s (1986) method has become one of the most often used methods for computing the fractal dimension of remotely sensed images (Lam and De Cola, 1993; De Jong and Burrough, 1995; Qiu et al., 1999; Lam et al., 2002; Myint, 2003). The purpose of this study is to expand Clarke’s 1986 work. Specifically, following Clarke’s (1986) triangular prism concept, we first propose three new methods to construct the triangular prisms used to compute the fractal dimension of digital images. The performance of the proposed methods and Clarke’s (1986) method is then tested on 45 synthetic surfaces and 72 real images. Photogrammetric Engineering & Remote Sensing Vol. 72, No. 4, April 2006, pp. 373–382. Department of Geography and Environmental Resources, Southern Illinois University-Carbondale, Carbondale, IL 62901 ([email protected]). PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING 0099-1112/06/7204–0373/$3.00/0 © 2006 American Society for Photogrammetry and Remote Sensing April 2006 373 04-130 3/14/06 9:05 PM Page 374 The Proposed Methods Clarke’s (1986) original triangular prism method makes use of a discrete representation of the elevations of the Earth’s surface such as in a digital elevation model (DEM). Based on this data structure, the method takes elevation values (or the equivalent of DN in an image) at the corners of squares, interpolates a center value, divides the square into four triangles, and then computes the top surface areas of the prisms which result from raising the triangles to their given elevations (Figure 1a). (Refer to Clarke, 1986 for a more complete discussion.) A critical step in the triangular prism method is the calculation of the top surface area of the prisms. Clarke’s (1986) method computes the top surface area by adding up the areas of four triangles formed by the corner pixels of a square and the interpolated center of the square. Obviously, Figure 1. (a) 3D view of the triangular prism method (after Clarke, 1986); (b) Top view of the corner pixels (a, b, c, and d ) and the center point (e) used in Clarke’s (1986) method (an example with step size 4). 374 April 2006 the ways in which the prisms are constructed affect the computed areas of the top surface, which in turn affect the computed D values of the surface. Critical to the triangular prism method is therefore the choice of the pixels used to construct the prisms. The three methods proposed in this study are inspired by approaches to construct the prisms that are different from Clarke’s (1986) method. Clarke (1986) used the average of the four elevations at the corners of a square to represent the center of the square. This approach clearly has the advantage of simple computation, but it may lead to error, because the interpolated value assigned to the center of the square may not lie on the surface being analyzed. This is especially the case where sharp tonal variation exists in the neighboring pixels. For example, suppose the DN values of the corner pixels are 1, 3, 2, 6, and the actual DN of the central pixel is 8. Using Clarke’s (1986) method, the interpolated value assigned to the central pixel would be 3, which is significantly different from its true value. In the three methods developed in this study, we propose to use the actual DN of the central pixel. The advantage of this approach is that every point used to calculate the area of the top surface can be assured to lie on the surface. In addition to using actual DN of the central pixel, the three proposed methods focus on the choice of edge pixels used to construct the triangular prisms. An edge pixel is defined in this paper as a pixel that is located on the edges of a square. Pixels a, b, c, and d in Figure 1b are examples of edge pixels. For a square of side length (or step size) , there are edge pixels on each edge of the square, and the total number of edge pixels is 4. As described above, Clarke’s (1986) method uses the corner pixels of a square and the interpolated center point to form the prisms. Two questions can be raised here. The first is how well the corner pixels actually represent the surface being analyzed. In other words, can approaches using other pixels better represent the surface. Our first two methods are inspired by this consideration. The second question to be raised is: can use of more edge pixels to approximate a surface yield more accurate estimates of D. Our third method attempts to provide answers to this question. The Max-Difference Method Our first method, which we call the Max-Difference method, seeks to reduce potential errors introduced by the use of corner pixels. It seems that in Clarke’s (1986) method, the corner pixels are chosen simply because of their physical locations. In other words, the actual DN values of edge pixels have no effect on the selection of pixels used to construct the prisms. The Max-Difference method represents a different approach to construct the prisms by taking into account the actual DN values of all edge pixels. Specifically, the MaxDifference method begins by searching a pixel on each edge of a square whose DN has the largest deviation from the central pixel. This procedure is reiterated for a certain number of geometrically increasing square sizes (e.g., 2, 4, 8, 16 . . . 2n) specified by the user. The four edge pixels, one on each edge, that meet this criterion are called max-difference edge pixels and used to construct the prisms. It should be clear that the max-difference edge pixels may or may not be corner pixels. In cases where the max-difference edge pixels are not located at the corners of a square, use of only corner pixels fails to capture the details of the surface raised by the max-difference edge pixels. By first searching the max-difference edge pixels, the Max-Difference method can avoid this problem and hence gives a better approximation to the surface. Computer implementation of the Max-Difference method is slightly more complex than that of Clarke’s (1986) method PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING 04-130 3/14/06 9:05 PM Page 375 due to the fact that the max-difference edge pixels can be any pixels on the edges of a square and, as a result, more base distances need to be calculated. The term “base distance” is used in this paper to mean the distance between two pixels on the base plane of the image (see Figure 1a). Furthermore, we use Tc to denote the base distance between the central pixel and an edge pixel and Te, the base distance between two edge pixels. In Figure 2 ea, eb, ec, and ed are examples of Tc, and ab, bc, cd, and da are examples of Te. In Clarke’s (1986) method, the base distances Tc between the center of a square and each of the corner pixels are the same and equal to d ( 12/2), and the base distances Te between each pair of the corner pixels are also the same and equal to , where is step size (Figure 1b). In other words, in Clarke’s (1986) method, only two sets of base distances need to be calculated. In our method, on the other hand, the base distances Tc between the central pixel and each of the max-difference edge pixels can be different. This is also true of the base distances Te between each pair of the max-difference edge pixels. As a result of this, each of these eight base distances needs to be computed. The Max-Difference method is implemented in six steps: Step 1. For the upper edge of a square, find the max-difference edge pixel (e.g., a in Figure 2) and record its location as La, La [1,], step size, Using the same procedure, find the max-difference edge pixels on the other three edges (e.g., b, c, d in Figure 2) and record their locations as Lb, Lc, and Ld, respectively, Lb, Lc, and Ld [1,], step size. Step 2. Calculate the base distances Tc between the central pixel and each of the four max-difference edge pixels. Equation 1 shows an example of how to compute Tc between e and a in Figure 2: Tc(ea) 1(La 1 d/2)2 (d/2)2 where La is the location of the max-difference edge pixel a, and is step size. Figure 2. Labels of edge pixels used in searching the max-difference edge pixels (a, b, c, and d) and computing the base distances (Tc) between the central pixel (e) and the four max-difference edge pixels, ea, eb, ec, and ed (an example with step size 4). PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING (1) Step 3. Calculate the base distances Te between each pair of the max-difference edge pixels. To calculate Te, we first define two arrays x(k) and y(k), where k [1,4] and step size, to record the x, y coordinates of the edge pixels in a raster coordinate system (see Figure 3). Equations 2-1 through 2-4 are used to compute the x, y coordinates of the pixels along the upper, right, lower, and left edges. upper x(i) i y(i) 1 lower right (2-1) x(i d) d 1 y(i d) i (2-2) left x(i 2d) d i 2 x(i 3d) 1 y(i 2d) d 1 (2-3) y(i 3d) d i 2 (2-4) where: i [1,] and is step size. Once the x, y coordinates of each edge pixel are determined, Te between each pair of the max-difference edge pixels can be calculated using the Pythagorean Theorem. Step 4. Calculate the top surface area of the prisms circumscribed by a square. The base distances Tc and Te obtained in steps 2-3, together with the DN values of the central pixel, the max-difference edge pixels and the corner pixels, are used to solve the lengths of sides of the triangles at the top of the prisms by the Pythagorean Theorem. The top surface area of each prism then can be computed using Heron’s formula (see Clarke, 1986 for the formulas used in this step). Note that the number of prisms used in the computation may vary, depending on the location of the maxdifference pixels. Adding up the top surface areas of all prisms gives the top surface area of the prisms circumscribed by the square. Figure 3. Labels of edge pixels used in computing the base distances (Te) between pairs of max-difference edge pixels, ab, bc, cd, and da, in a raster coordinate system (an example with step size 4). April 2006 375 04-130 3/14/06 9:05 PM Page 376 Step 5. Calculate the total area of the top surface by summing up the top surface areas of the prisms circumscribed by all squares of a given step size needed to cover the surface (Figure 4). Step 6. Compute the fractal dimension (D) value. The total area of the top surface can be calculated repeatedly for increasing step sizes by iterating Steps 1 through 5. To obtain equally-spaced observations on the independent variable for the regression, step size is increased by powers of 2 (i.e., 2, 4, 8, 16 . . . 2n). As the size of the squares used to cover the surface increases, more details in the top surface get lost and, as a result, the estimated total area of the top surface decreases (Figure 4). Once the total areas of the top surface are calculated for a certain number of step sizes, plot log (total top surface area) against log (step size), fit a least-squares regression line through the data points, and calculate the slope (b) of the regression line. The fractal dimension (D) of the surface then is computed as D 2 b, where b is the slope of the log-log regression line. The procedure described above can be illustrated with the following example (Figure 4). Suppose we use four step sizes (i.e., 2, 4, 8, and 16) to estimate the D of an image. We first compute the total area of the top surface for step size 2, resulting in an estimate of the top surface area (e.g., 3259.6). We repeat this calculation for step sizes 4, 8, and 16, and obtain another three sets of estimated area of the top surface (e.g., 3196.1, 3158.1, and 3110.2). Plot log (3259.6) versus log (2), log (3196.1) versus log (4), and so forth; fit a least-squares regression line through the four data points and calculate the slope (b) of the regression line (e.g., b 0.0296). The fractal dimension (D) of the image then is computed as D 2 b 2 (0.0296) 2.0296. The Mean-Difference Method Our second method, which we call the Mean-Difference method, is similar to the Max-Difference method in that, in choosing the pixels used to construct the prisms, it takes into account the actual DN values of all edge pixels. The difference is in the criterion used. Specifically, the MeanDifference method uses a pixel on each edge of a square whose DN deviation from the central pixel is closest to the mean of DN deviations from the central pixel to all pixels on the same edge. The pixels that meet this criterion are called mean-difference edge pixels and used to construct the prisms. Similarly to the Max-Difference method, computer implementation of the Mean-Difference method also consists of six steps. The procedures and equations used in Steps 2 through 6 are the same as in the Max-Difference method described above. The only difference is in the first step. Specifically, the first step in the Mean-Difference method involves searching a mean-difference pixel on each edge of a square using the following procedure. For each edge of a square (e.g., the upper edge), find the mean-difference edge pixel (e.g., a in Figure 2) using Equations 3-1 through 3-3 and record its location as La, La [1,], step size. h(i) ABS [h(e) h(i)] (3-1) h MEAN [h(i)] (3-2) MeanPixel MIN {ABS[ h h(i)]} (3-3) where: i ith pixel on an edge, i [1,], is step size; h(i) absolute DN difference between the central pixel and ith edge pixel; h(e) DN of the central pixel; h(i) DN of ith edge pixel; h mean of DN deviations from the central pixel to all pixels on the same edge; ABS, MEAN, MIN the absolute, mean, and minimum function, respectively; and MeanPixel the mean-difference pixel. Using the same procedure, finding the mean-difference edge pixels on the other three edges (e.g., b, c, and d in Figure 2) and recording their locations as Lb, Lc, and Ld, respectively, Lb, Lc, and Ld [1,], step size. Figure 4. Illustration of the reiteration procedure to establish the relationship between the top surface area of the prisms and the step size (or square size) used in the triangular prism method: (a) step size 2, top surface area (64 squares); (b) step size 4, top surface area (16 squares); (c) step size 8, top surface area (4 squares); (d) step size 16, top surface area (1 square); (e) fitting a least-squares regression line to derive fractal dimension (D). 376 April 2006 The Eight-Pixel Method Our third method, which we call the Eight-Pixel method, is inspired by the question whether use of more pixels can better approximate an image surface. The idea is that, except for step size 1, there are more than four pixels on the edges of a square available to represent the surface. More importantly, as the size of the square used to cover the surface increases, it may well be the case that the geometry of the top surface circumscribed by the square becomes increasingly “complicated,” and use of only four pixels may not represent the surface so well. As such, it seems reasonable to argue that using more pixels to represent a surface should be able to capture more details of the geometry of PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING 04-130 3/14/06 9:05 PM Page 377 used. Several researchers (e.g., Tate, 1998; Lam et al., 2002) have suggested that a good strategy for evaluating fractal computation methods is to test them using synthetic data with known generated D values. In this approach, simulated data can serve as a benchmark for assessing the performance of the computation methods used. A comparison between computed and generated D values provides an indication of the performance of the methods being examined. Fractional Brownian motion (FBM) is a widely used method of generating fractal surfaces with known D values. The variogram of an FBM surface has the form (Goodchild, 1980; Goodchild and Mark, 1987): E[z(x) z(x d)]2 k(|d|)2H Figure 5. Top view of the eight edge pixels (a, b, c, d, e, f, g, and h) and the central pixel (o) used in the Eight-Pixel method (an example with step size 4). the surface. This is particularly true when there is sharp gray-level variation in the neighboring pixels. Based on this line of reasoning, the Eight-Pixel method attempts to better approximate a surface by using eight edge pixels. The eight pixels used to construct the prisms are the four corner pixels as in Clarke’s (1986) method, and four middle pixels on the edges of a square. A middle pixel is defined as a pixel that is halfway between a pair of corner pixels (Figure 5). The computation of the top surface area of the eight prisms in the Eight-Pixel method is similar to the procedures described in Clarke (1986). In terms of computer implementation, three differences exist between these two methods. • • • First, since the Eight-Pixel method uses eight edge pixels, two sets of base distances between the central pixel and the edge pixels (Tc) need to be calculated. The first is the base distances between the central pixel and each of the corner pixels (e.g., oa, oc, oe, and og in Figure 5). This set of base distances are calculated in the same way as in Clarke’s (1986) method, and they are all equal to d ( 12/2), where is step size. Another set of base distances needed to be calculated is the distances between the central pixel and each of the four middle pixels (e.g., ob, od, of, and oh in Figure 5), which have the same value of 0.5. Second, in the Eight-Pixel method, the base distances between two neighboring edge pixels (Te) (e.g., ab, bc, cd, and de in Figure 5) are the same and equal to 0.5, whereas this value is in Clarke’s (1986) method. Third, eight triangles are used in the Eight-Pixel method to compute the top surface area of the prisms, whereas four triangles are used in Clarke’s (1986) method. (4) Where E[ ] denotes the statistical expectation, z(x) is the elevation of the surface at coordinates denoted by the vector x, d is a displacement vector, k is a constant, |d| is the magnitude (length) of the displacement vector, and H is a parameter in the range 0 to 1. The fractal dimensions (D) of simulated surfaces are equal to D 3 H (Goodchild, 1980). The FBM surfaces used in this study were generated using the shear displacement method (Saupe, 1988). Detailed description of this method can be found in Goodchild (1980) and Lam and De Cola (1993). Since synthetic surfaces are realizations of a stochastic process, it is necessary to use multiple sample data sets at each D value to adequately compare the performance of different methods of computing fractal dimensions. In this study, five samples of synthetic surfaces were generated for each H level (or each D value) from 0.1 through 0.9 in steps of 0.1 (i.e., H 0.1, 0.2, . . . 0.9 or D 2.9, 2.8, . . . 2.1). In other words, a total of 45 synthetic surfaces were used in this study. Following the approach presented in Tate (1998) and Lam et al. (2002), an identical seed value of the random number generator was used to generate each set of the surfaces with varying H. For each surface generated, a total of 3,000 fault lines were applied. Each of the synthetic surfaces was generated using a 513 513 grid. The generated surfaces were stretched to 8-bit images with DN ranging from 0 to 255. Three sample surfaces and their corresponding images representing generated D 2.1, 2.5, and 2.9 are displayed in Figure 6. To demonstrate the utility of fractal techniques in the analysis of real remote sensing images, we also tested the four computation methods on 72 real images. The real images used for this purpose were extracted from a digital aerial photograph of Carbondale, Illinois acquired on 25 January 2001. The digital aerial photograph has a spatial resolution of 0.9843 ft 0.9843 ft and was acquired in three visible bands: red, green, and blue. Six land-use/land-cover types with different textural appearance, i.e., crop field, forest, parking lot, pasture, residential area, and lake surface, were selected for this analysis. For each of the six landuse/land-cover types, 12 samples of 129 by 129 pixels in the red band were used in the computation. The 72 sample images used in this analysis are displayed in Figure 7. Note that each column in Figure 7 contains a different sample of each land-use/land-cover type (labeled crop field, forest, parking lot, pasture, residential, and lake surface) while each row consists of 12 different samples (labeled S1 to S12) of the same land-use/land-cover type. Test Data The estimates of D computed from real data such as DEMs and remote sensing images are of unknown accuracy. As such, computed D values of real data may not explain precisely the effectiveness of the computation methods PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING Results and Discussion The input parameters used in the implementation of the proposed methods and Clarke’s (1986) method are specified April 2006 377 04-130 3/14/06 9:05 PM Page 378 Figure 6. Samples of synthetic surfaces and their corresponding images representing generated D 2.1 (a), 2.5 (b), and 2.9 (c). The upper-left corner of the images (right column) corresponds to the origin (0,0) of the synthetic surfaces (left column). 378 April 2006 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING 04-130 3/14/06 9:05 PM Page 379 Figure 7. Samples of real images: each column contains a different sample of each land-use/land-cover type (labeled crop field, forest, parking lot, pasture, residential, and lake surface); each row consists of 12 different samples (labeled S1 to S12) of the same land-use/land-cover type. as follows. Because the proposed methods use actual DN of the central pixel, the starting step size used in these methods was 2, followed by step sizes of 4, 8, 16 . . . , and up to 512. In other words, nine step sizes were used in the proposed methods. Clarke’s (1986) method does not require the use of actual DN of the central pixel. Therefore, the starting step size used in this method was 1 and the number of step sizes used was ten. To compute the D values of the real images, seven step sizes were used in the proposed methods and eight in Clarke’s (1986) method. Following the work of Lam et al. (2002), all the methods tested in this study, including Clarke’s (1986) method, used step size () instead of step size squared (2) in the regression to derive the slope of the regression line. As such, the term “Clarke’s method” used in the following text should be interpreted as a modified version of Clarke’s 1986 method. For each generated D value, the three proposed methods and Clarke’s (1986) method were used to compute five estimates of D using the synthetic images described above. The mean of the computed D values obtained by each of the four methods at each generated D level then was calculated and the results are presented in Figure 8. The deviations of the computed D values from generated D values at each D level are plotted in Figure 9. The standard deviation and root mean squared error (RMSE) of the computed D values are reported in Table 1. RMSE were calculated using the n following formula: RMSE e a [Dest(i) Dgen(i)]2 fnn, where B i1 Dest is computed D, Dgen is generated D, and n is the number of synthetic surfaces used. It can be seen from Figures 8 and 9 that the four computation methods returned smaller estimated D values than the generated D values for the synthetic images over the entire range of fractal dimension. Measured by the extent of the deviations of the estimated D values from the generated D values, the performance of the four methods varies with respect to the smoothness or roughness of the images used. Specifically, the four methods produced D values PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING Figure 8. Plot of mean computed fractal dimensions obtained using Clarke’s (1986), Max-Difference, MeanDifference, and Eight-Pixel methods versus generated fractal dimensions. closer to those generated ones for images with smooth to quite “rough” surfaces (e.g., generated D 2.1–2.2 and 2.5–2.7), but they resulted in larger deviations for extremely “rough” images (e.g., generated D 2.9). April 2006 379 04-130 3/14/06 9:05 PM Page 380 Figure 9. Deviations of computed fractal dimensions from generated fractal dimensions resulted from Clarke’s (1986), Max-Difference, Mean-Difference, and Eight-Pixel methods. Another pattern displayed in Figures 8 and 9 is that, in terms of the deviations of computed D values from generated D values, the three proposed methods appear to generally perform better than Clarke’s (1986) method. This also is reflected in the generally smaller RMSE yielded by the three proposed methods compared with those generated by Clarke’s (1986) method (Table 1). Furthermore, the results from the test for equality of means show that the differences between the mean computed D values obtained by the three proposed methods and those obtained by Clarke’s (1986) method are all statistically significant where generated D 2.4 (see Table 1). Among the three proposed method, the Max-Difference method returned improved estimates of D over the entire range of fractal dimension and the improvements were statistically significant. This result seems to suggest that use of max-difference edge pixels appears able to capture more details of the geometry of image surfaces. TABLE 1. COMPARISON Generated D Clarke stdv OF THE PERFORMANCE 2.1 0.0019 0.0850 Max-Difference stdv 0.0033 0.0751 RMSE delta mean 0.0099*** t (6.087) Mean-Difference stdv 0.0018 RMSE 0.0789 delta mean 0.0060*** t (4.979) Eight-Pixel stdv 0.0020 0.0788 RMSE delta mean 0.0061*** t (5.049) RMSE OF The improved results achieved by the Mean-Difference method for images with complex textures suggest that use of mean-difference edge pixels also appears to be a sound approach to approximate image surfaces. The Eight-Pixel method was outperformed by the MaxDifference method over the range of D 2.1–2.5. Nevertheless, compared to Clarke’s (1986) method, the Eight-Pixel method also delivered improved estimates of D and the improvements were statistically significant for relatively “rough” images with generated D 2.4. These results seem to suggest that the strategy of using more edge pixels to approximate an image surface appears generally effective. This seems especially true of the images with greater textural complexity. This finding makes sense because using more pixels that lie on the image surface should be able to capture more details of the geometry of the surface. Another factor that may have contributed to the improved accuracies of computed D values using the proposed methods is their use of actual DN values of the central pixel. The advantage of using actual DN values of the central pixel is that it ensures that all the pixels used to construct the prisms lie on the image surface being analyzed and therefore can better approximate the surface. A disadvantage of this approach is the loss of one data point for the generation of the regression line. The results from the analyses of the 72 real images are shown in Figure 10. Three observations can be made here. First, the three proposed methods and Clarke’s (1986) method appear generally effective in differentiating the six landuse/land-cover types examined in this study. The four methods did a particularly good job in distinguishing those land-cover features with apparently different textures such as parking lot, residential area, lake surface, forest, and crop field. This is reflected in the relatively large “distances” between the computed D values of these land-use/land-cover types. Second, for the same data set, that is, the same image samples of a land-use/land-cover type, the computed D values obtained by the four methods vary. Take the land-use type “residential area” for example. The mean computed D values for the “residential” image data set obtained using Clarke’s (1986), Max-Difference, Mean-Difference, and Eight-Pixel methods are 2.506, 2.497, 2.566, and 2.494, respectively. CLARKE’ (1986), MAX-DIFFERENCE, MEAN-DIFFERENCE, USING THE SYNTHETIC SURFACES 2.2 2.3 2.4 0.0056 0.1693 0.0076 0.1599 0.0095* (2.208) 0.0059 0.1624 0.0069* (1.887) 0.0059 0.1626 0.0067 (1.845) 0.0155 0.2082 0.0177 0.1889 0.0196* (1.873) 0.0155 0.1945 0.0137 (1.392) 0.0170 0.1946 0.0138 (1.345) 0.0230 0.1994 0.0230 0.1614 0.0382** (2.638) 0.0236 0.1713 0.0284* (1.883) 0.0249 0.1704 0.0294* (1.940) 2.5 2.6 2.7 AND EIGHT-PIXEL METHODS 2.8 2.9 0.0284 0.0251 0.0131 0.0076 0.0086 0.1842 0.1767 0.1862 0.2344 0.3150 0.0223 0.0171 0.0071 0.0103 0.0063 0.1310 0.1121 0.1112 0.1427 0.2142 0.0529** 0.0643*** 0.0749*** 0.0919*** 0.1008*** (3.287) (4.749) (11.209) (16.386) (21.248) 0.0286 0.0250 0.0136 0.0089 0.0092 0.1396 0.1155 0.1096 0.1454 0.2191 0.0453** 0.0621*** 0.0769*** 0.0890*** 0.0959*** (2.515) (3.894) (9.282) (17.411) (17.202) 0.0293 0.0253 0.0137 0.0074 0.0091 0.137 0.1109 0.1033 0.1388 0.2121 0.0480** 0.0668*** 0.0833*** 0.0956*** 0.1030*** (2.631) (4.188) (9.816) (20.088) (18.364) All Images 0.2036 0.1495 0.1541 0.1511 stdv standard deviation of computed D values. RMSE root mean sqaured error of computed D values. delta mean (mean computed D values obtained by the proposed method) minus (mean computed D values obtained by Clarke’s method). t t values from the test for equality of mean computed D values obtained by the proposed method and Clarke’s method. ***p 0.01; **p 0.05; *p 0.1; two tailed tests. 380 April 2006 PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING 04-130 3/14/06 9:05 PM Page 381 Figure 10. Computed fractal dimensions of the real images representing six land-use/land-cover types using Clarke’s (1986), Max-Difference, Mean-Difference, and Eight-Pixel methods. Third, when applied to the images representing relatively homogeneous land-cover features such as pasture and crop fields, the four methods returned consistent estimates of D, reflected in the relatively small variations in the estimated D values for these land-cover features. When applied to the images with less homogeneous land-use/land-cover features such as parking lots, residential areas and, to a lesser extent, forest, on the other hand, the variations in the estimated D values for the same land-use/land-cover type appear relatively large. This result may suggest that the spatial arrangements of individual objects in these images, such as the orientation, size and spacing of houses, cars, and trees, may all have had an effect on the computed D values. The apparently large variations in the computed D values for the images representing “lake surface” seemed counter-intuitive since “lake surface” is usually perceived “smooth.” A possible explanation to this seemingly unusual result is that the aerial photograph used in this study was taken in January, a time when parts of the lakes in the study area were frozen. As such, factors such as the size, depth, and distribution of ice in the lake were highly variable from one image to another. All these factors appear to have affected the spatial patterns of tonal variations (i.e., brightness or darkness) in the images used, which in turn resulted in large variations in the computed D values. It should be noted that, because the proposed methods involve comparing the DN values of all edge pixels and using more edge pixels, they are computationally more complex than Clarke’s (1986) method. For example, the computing time needed to calculate the D value of a real image of 129 129 pixels with complex textures (estimated D 2.61 and 2.69) was about eight seconds using the Max-difference method, compared to about two seconds using Clarke’s (1986) method. The findings of the present study have implications for the application of fractal techniques to remote sensing image analysis. First, our results from the analyses of the 72 real images suggest that fractal techniques appear to be a promising tool for characterizing the textural complexity of remotely sensed images. The computed D values provide a potentially useful source of textural information that could be used to supplement digital image classification problems. Since the main purpose of the present study was to develop new PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING algorithms and compare the performance of different computation methods based on the triangular prism concept, we only calculated a single global D value, that is, an average (or mean) D, for each synthetic and real image. This approach has been used in several studies focusing on algorithm evaluation, such as in Tate (1998) and Lam et al. (2002). It should be noted that when fractal dimensions are to be used as textural information to segment remote sensing images, computing global D values does not serve the purpose. What is needed in these applications is the “local” D values reflecting local tonal variations of different land-use/land-cover types. Such local D values can be computed using local “moving window” techniques. Although a thorough discussion of the implementation of the triangular prism method at local window levels is beyond the scope of this study, we direct readers interested in this issue to the work presented by De Jong and Burrough (1995). Second, our results indicate that the triangular prism technique, as implemented in our proposed methods and Clarke’s (1986) method, appears more suitable for computing the fractal dimension of remote sensing images with smooth to quite rough surfaces, but it tends to perform poorly on extremely rough images. This means that, to obtain reliable results using a fractal technique such as the triangular prism, one would need some knowledge about the textural complexity of the image to be analyzed. Fortunately, a large number of other techniques, such as the co-occurrence matrices (Haralick et al., 1973), local variance (Woodcock and Strahler, 1987), wavelets (Mallat, 1989), and spatial autocorrelation statistics (Cliff and Ord, 1973), are available for characterizing textural features in digital images. As such, using other texture analysis techniques to develop a “feel” for the image data set to be analyzed appears desirable before a fractal dimension computation method is applied. Third, our study also shows that even using the same approach for computing fractal dimension, such as the triangular prism, different estimation procedures, such as the four implementations of the triangular prism approach examined in this study, may yield different computed D values for the same data set. Studies comparing different approaches for computing fractal dimension (e.g., the variogram, triangular prism, and isarithm) have also reported that variation in computed D values can be introduced by the choice of general approaches (Tate, 1998; Lam et al., 2002; Myint, 2003). Taken together, these results seem to suggest that the choice of computation methods is an important issue in the application of fractal techniques to digital image analysis. Researchers need to be aware of the comparative performance of different computation methods proposed in the literature and the biases that are associated with a particular method or a particular implementation procedure. Conclusions How to extract the often complex and erratic textures in digital images and use such information to improve image classification has been a major research issue for remote sensing scholars for decades. In this context, fractal geometry appears appealing because it offers us powerful tools for analyzing fragmented and irregular forms, which seem characteristic of textural features in remotely sensed images. There is a large and growing literature examining the potential of the fractal model in digital image processing. A review of this literature suggests that fractal techniques appear capable of capturing certain aspect of the spatial variations of gray-levels in digital images that may not be captured by other texture analysis techniques (Mandelbrot, 1977; Pentland, 1984). As such, incorporating fractal parameters such as the fractal dimension April 2006 381 04-130 3/14/06 9:05 PM Page 382 into a digital image processing procedure appears to be a promising approach to enhancing image analysis results. It should be noted, however, that the fractal dimension describes only one aspect (i.e., the shape) of the textures in digital images. This means that use of fractal dimension alone may not be sufficient to completely characterize image textures. It appears that the utility of fractal dimension may be explored to a fuller extent when it is used in conjunction with other texture measures and perhaps spectral information as well. Developing efficient and reliable algorithms for computing fractal dimension is key to fully exploring the potential of fractal techniques in remote sensing applications. In this study, we proposed three new alternatives based on Clarke’s (1986) triangular prism concept to compute the fractal dimension of remotely sensed images. Our results indicate that in comparison with Clarke’s (1986) method, the proposed methods appear able to generate improved estimates of fractal dimensions for synthetic images with complex textures. Results from the analysis of real remote sensing images show that the proposed methods appear effective in differentiating different land-use/land-cover types examined in this study. The performance of the three proposed methods and Clarke’s (1986) method reported in this study was based on the analyses of 45 synthetic surfaces and 72 real images. More thorough evaluation of the proposed methods clearly is needed to establish the merits and drawbacks of these new methods in comparison with other existing methods. In this context, testing the proposed methods on a variety of real image data and developing efficient algorithms to compute local D values of real images, analogous to the work of De Jong and Burrough (1995), are especially desirable because the utility of any computation methods is ultimately measured by their performance in the analysis of real remote sensing images. It is our hope that the results presented in this paper could spark further interest in the evaluation of various fractal computation methods and, thereby, contribute to the development of more effective techniques and efficient algorithms for image texture analysis. Acknowledgments I am grateful to the office of Research Development and Administration (ORDA) at Southern Illinois University carbondale for providing support for this research. I would like to thank the three anonymous reviewers for their very helpful comments on an earlier draft of this paper. References Burrough, P.A., 1981. Fractal dimensions of landscapes and other environmental data, Nature, 294(19):240–242. Clarke, K.C., 1986. Computation of the fractal dimension of topographic surfaces using the triangular prism surface area method, Computers and Geosciences, 12(5):713–722. Cliff, A.D., and J.K. Ord, 1973. Spatial Autocorrelation, Pion Limited, London. De Cola, L., 1989. Fractal analysis of a classified Landsat scene, Photogrammetric Engineering & Remote Sensing, 55(5):601–610. De Jong, S.M., and P.A. Burrough, 1995. A fractal approach to the classification of Mediterranean vegetation types in remotely 382 April 2006 sensed images, Photogrammetric Engineering & Remote Sensing, 61(8):1041–1053. Gong, P., and P.J. Howarth, 1990. The use of structural information for improving land-cover classification accuracies at the ruralurban fringe, Photogrammetric Engineering & Remote Sensing, 56(1):67–73. Goodchild, M.F., 1980. Fractals and the accuracy of geographical measures, Mathematical Geology, 12(2):85–98. Goodchild, M.F., and D.M. Mark, 1987. The fractal nature of geographic phenomena, Annals of the Association of American Geographers, 77(2):265–278. Haralick, R.M., K. Shanmugan, and I. Dinstein, 1973. Texture features for image classification, IEEE Transactions on Systems, Man, and Cybernetics, SMC-3(6):610–621. Klinkenberg, B., and M.F. Goodchild, 1992. The fractal properties of topography: A comparison of methods, Earth Surface Processes and Landforms, 17:217–234. Lam, N.S.-N., 1990. Description and measurement of Landsat TM images using fractals, Photogrammetric Engineering & Remote Sensing, 56(2):187–195. Lam, N.S.-N., and L. De Cola, 1993. Fractal simulation and interpolation, Fractals in Geography (N.S.-N. Lam and L. De Cola, editors), Prentice Hall, New Jersey, pp. 56–74. Lam, N.S.-N., H.L. Qiu, D.A. Quattrochi, and C.W. Emerson, 2002. An evaluation of fractal methods for characterizing image complexity, Cartography and Geographic Information Science, 29(1):25–35. Mallat, S.G., 1989. A theory for multi-resolution signal decomposition: The wavelet representation, IEEE Transactions on Pattern Analysis and Machine Intelligence, 11(7):674–693. Mandelbrot, B.B., 1977. Fractals: Form, Chance and Dimension, W.H. Freeman and Company, San Francisco, California, 265 p. Mark, D.M., and P.B. Aronson, 1984. Scale-dependent fractal dimensions of topographic surfaces, Mathematical Geology, 16(7):671–683. Myint, S.W., 2003. Fractal approaches in texture analysis and classification of remotely sensed data: Comparisons with spatial autocorrelation techniques and simple descriptive statistics, International Journal of Remote Sensing, 24(9):1925–1947. Pentland, A.P., 1984. Fractal-based descriptions of natural scenes, IEEE Transactions on Pattern Analysis and Machine Intelligence, 6(6):661–674. Pratt, W.K., O.D. Faugeras, and A. Gagalowicz, 1978. Visual discrimination of stochastic texture fields, IEEE Transactions on Systems, Man and Cybernetics, 8(11):796–804. Roy, A.G., G. Gravel, and C. Gauthier, 1987. Measuring the dimension of surfaces: A review and appraisal of different methods, Proceedings of the Eighth International Symposium on Computer-Assisted Cartography (Auto-Carto 8), pp. 68–77. Saupe, D., 1988. Algorithms for random fractals, The Science of Fractal Images (H.O. Peitgen and D. Saupe, editors), SpringerVerlag, New York, New York, pp. 71–136. Shelberg, M.C., N.S.-N. Lam, and H. Moellering, 1983. Measuring the fractal dimension of surfaces, Proceedings of the Sixth International Symposium on Computer-Assisted Cartography (Auto-Carto 6), pp. 319–328. Tate, N.J., 1998. Estimating the fractal dimension of synthetic topographic surfaces, Computers and Geosciences, 24(4):325–334. Woodcock, C.E., and A.H. Strahler, 1987. The factor of scale in remote sensing, Remote Sensing of Environment, 21:311–332. (Received 02 November 2004; accepted 22 February 2005; revised 21 March 2005) PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
© Copyright 2026 Paperzz