GIS for Coastal Zone Management - Chapter 7

CHAPTER SEVEN Exploring the Optimum Spatial Resolution for Satellite Imagery: A Coastal Area Case Study Chul-sue Hwang and Cha Yong Ku 7.1 INTRODUCTION Researchers in geography are much interested in issues of scale and resolution, and the strong influence these may have on the results of analyses of spatial patterns (Harvey, 1968; Stone, 1972; Goodchild and Quattrochi, 1997). Issues of scale usually arise from not knowing the optimal unit size for a study area; while MAUP (the Modifiable Area Unit Problem), one of the fundamental research problems in geographical information science, can also be considered to be related to the understanding of the scale for spatial data (Openshaw, 1984; Fotheringham and Wong, 1991). GIS and remote sensing have shown active interests in both the operational scale and measurement scale of data. Technological developments in GIS and remote sensing have led to the gradual detailing and elaboration of the measurement scale. In the field of remote sensing, for example, the 78m resolution Multi-Spectral Scanner (MSS) data of the 1970s has given way to 30m Thematic Mapper (TM) data in the 1980s, SPOT in the 1990s, and 1m IKONOS imagery in the 21st century. Such improvements in the measurement scale will allow for the new understanding of geographical phenomena that could not be measured at the previous scales. Changes of the ground resolution of satellite imagery have modified the inherent characteristics of imagery due to the "scale effect" (Goodchild and Quattrochi, 1997). Therefore, researchers have to select the image with the optimum resolution among the satellite imagery at diverse resolutions, at which a specific geographical phenomenon can be understood more effectively. In order to explore the optimum resolution, the objective procedures or indices must be designed for the image characteristics at different resolutions. They should be able to represent the changes in image characteristics that occur at different resolutions. The aim of the present study is to propose a procedure for identifying the optimum resolution for the study of coastal wetlands. These wetlands, as the point where the ecosystem of the land meets the ecosystem of the ocean, have unique © 2005 by CRC Press LLC and diverse topographical features, giving rise to the formation of vegetation colonies, and exhibits diverse spatial distribution due to diverse environmental factors. Due to the dense vegetation and unstable topography of the coastal wetland, it is more effective to use satellite imagery to understand its characteristics, rather than through field investigation (Lyon and McCarthy, 1995). However, it is not necessarily effective to use the satellite imagery of higher ground resolution in the coastal area. For example, in the case of the salt marsh where colonies appear as clusters, the use of low-resolution imagery of the approximate size of the clusters is more effective (McNairn et al., 1993). In this vein, this study will attempt to find procedures and techniques for the optimal spatial resolution of satellite imagery in the study of coastal wetlands. To achieve this goal, it firstly selected the indices considering the following conditions: (1) the sensitivity of the indices, which change in response to the different resolutions, should be high enough to respond accurately to the changes in image characteristics; (2) the indices should be the ones that can provide the approximate estimate of the final result only with the minimal procedure; (3) the indices should reflect the classification accuracy of the attribute information. We then coarsened the resolution of the satellite data, to produce images at various resolutions, and observed the changes in the textural and attribute information characteristics of these images at various resolutions. As a result, the study was able to decide the operational scale, an optimum resolution in the use of satellite imagery. 7.2 SATELLITE IMAGERY CHARACTERISTICS BY RESOLUTION 7.2.1 Textural Characteristics of Satellite Imagery The local variance method is frequently used in the search for the optimum resolution. Proposed by Woodcock and Strahler (1987), the local variance method measures the variance by placing a 3 x 3 moving window over the image data, to capture the textural characteristics of the image. However, local variance is limited by the fact that it relies on the overall variance of the image. In other words, the local variance measurements of one image cannot be compared to the results for another image. In addition, using this index makes it impossible to understand the spatial distribution of the variance within the image. Cao (1992), in an attempt to calculate the fractal dimension of satellite imagery, compared and analysed the Isarithm method (Goodchild, 1980), the Variogram method (Mark and Aronson, 1984) and the Triangular prism method (Clarke, 1986). From the analysis, he found the Isarithm method most useful. In general, the fractal dimension D of satellite imagery appears as a real number between the 2D and 3D. As the complexity of the image’s spatial structure increases, the value of the fractal dimension also gets higher. Within this principle, obtaining the fractal dimension gives a numerical representation of the satellite imagery characteristics by resolution. © 2005 by CRC Press LLC In order to understand how the distribution characteristics of images change during the coarsening of resolution, the images of variances as well as the images of averages are needed. It is similar logic that we use the variance to make up the weak points in using the arithmetic mean (or average) to represent a data set. The variance of an image set provides a measure of the average squared deviation of a set of values around the mean. While the images of averages represent the image as one pixel, taking the arithmetic mean of the original pixels, the images of variances represent the dispersion characteristics or the variability of the original image during the coarsening of the resolution (Figure 7.1). During the resolution coarsening, if the heterogeneity between the pixels is large, the variance of the generated image will be large. Conversely, if the pixels are of similar value, the variance of the generated image will be small. Low variances are recorded for smooth and wide regions during the resolution coarsening, and high variances are recorded for narrow and complex regions. x G x2 G ว images of averages xm 4 i 1 xi G 4 x3 G x4 G ว images of variancesG xv i41 (xmxi)2 G 41 Figure 7.1. Calculating images of averages and images of variances. Understanding the spatial distribution pattern of the images of variances derived from the process of the resolution transformation leads to the analysis of changes in image characteristics by resolutions. Therefore, the analysis of the spatial distribution of the images of variances can be used to understand the image transformation that follows the resolution coarsening. The spatial autocorrelation of the images of variances is selected to measure the spatial distribution in which the image characteristics transform. The spatial autocorrelation using Moran`s I is generally appropriate to measure the distribution pattern for data with a continuous variable such as satellite imagery. 7.2.2 Measuring the Spectral Separability of the Training Set Divergence is a common measurement to estimate the statistical separability of satellite imagery. Divergence is measured using the average of the statistics by class and the covariance matrix gathered through the supervised classification of the training set. However, since divergence has a tendency to increase abruptly as the distance between the classes increases, it is difficult to generalise divergence (Jensen, 1996). The Jeffreys-Matusita distance (Matusita, 1966) is a value calculated by the exponential transformation of the Bhattacharyya distance that has been modified from divergence (Kailath, 1967; Wacker, 1971). This J-M distance is more widely © 2005 by CRC Press LLC used than divergence for calculating the spectral separability of multi-variate classified classes because the J-M distance does not exhibit the abrupt increase in the spectral separability. However, complex calculations and a long processing time are some of the disadvantages of the J-M distance. Mahalanobis distance (Mahalanobis, 1936) is a value that takes into account both the correlation and distribution features between the data in terms of Euclidean distances. Mahalanobis distance multiplies the Euclidean distance by the covariance matrix between classes. In contrast to the J-M distance, Mahalanobis distance requires simpler calculations and less processing time. In addition, it does not exhibit the abrupt increases in the spectral separability. For these reasons, Mahalanobis distance is the most widely used method for measuring the spectral separability. 7.3 PROCEDURE FOR EXPLORING THE OPTIMUM SPATIAL RESOLUTION The optimum spatial resolution using satellite imagery is the one that provides good representation of the attribute information, and is referred to as the ‘operational scale.’ In order to select the image with the optimum resolution from satellite images at different resolutions, the objective procedures or indices must be designed for the image characteristics at different resolutions. Figure 7.2 The proposed procedure to explore the optimum spatial resolution of satellite imagery. © 2005 by CRC Press LLC However, it is unreasonable to rely on a single index to select the optimum spatial resolution. Indices have their own specific objectives to understand the characteristics of the imagery. Therefore, this study will attempt to combine the various indices and present a procedure for exploring the optimum spatial resolution (Figure 7.2). Until now, existing studies on selecting the optimum resolution have focused only on the textural changes of the images at different resolutions. Local variance and fractal methods are particularly mentioned in such studies. However, because they only reflect the result without illustrating processing procedure of attribute information obtained from the images, they should not be used to understand the changes in attribute information. This study will attempt to approach the exploration of the optimum resolution using satellite imagery using a two-step procedure. First, through the measurement of the image`s textural characteristics, the resolution of action will be selected, and then, through the estimation of the spectral separability, the adaptive resolution for the attribute information will be chosen. 7.3.1 Textural Characteristics: Searching for the Resolution of Action If the resolution of the original image is gradually scaled down to produce images at low resolutions, the image achieves a structure much like that of a pyramid. This method of the scale transformation is called the pyramid model (Figure 7.3). When producing an image at a low resolution, producing both images of averages and images of variances at the same time is more effective for the analysis of the satellite imagery. The images of variances are the ones that reflect the differences between the original images and the coarsened images during the resolution coarsening. Figure 7.3 Construction of an image pyramid (source: Richards, 1993, p. 33). x Step 1: Moran`s I is calculated based on the images of variances. It can be said that the higher the Moran`s I value, the higher the spatial autocorrelation will be and the spatial distribution pattern will then © 2005 by CRC Press LLC ... - tailieumienphi.vn