Enhancing Binar y Change Detection Performance Using A Moving Thr eshold Window (MTW) Appr oach

This study introduced a new concept, the Moving Threshold Window (MTW), for binary change detection. An automated MTW-based calibration model was developed and evaluated using a case study. The MTW-based model is free from the assumption of symmetry for difference and ratio types of change-enhanced images, unlike traditional binary change detection methods. The MTW-based calibration model outperformed the traditional binary change detection methods based on the Symmetric Threshold Window (STW) for both single and multiple change-enhanced images of the study area. In most of the calibrations, the optimum thresholds resulting in the highest Kappa coefficient were asymmetric. Three major factors may explain the asymmetric characteristics of the optimum thresholds, including: different atmospheric conditions found in the two dates of imagery, different look angles associated with the two dates of imagery, and the nature of the change information. Multiple change-enhanced images generally produced higher accuracies than single changeenhanced images using both the MTWand STW-based models. Introduction Monitoring the change of biophysical characteristics in ecosystems has been one of the major research topics in environmental science (e.g., Singh, 1989; Coppin and Bauer, 1996; Hayes and Sader, 2001; Lunetta et al., 2002; Chen et al., 2003; Lu et al., 2005). Changes in biophysical materials and human-made structures on the surface of Earth are quite dynamic and need to be accurately identified in a timely manner for management and planning purposes. Remote sensing provides valuable multiple-date data for change detection (Jensen, 2005). Remote sensing-based change detection has been generally used to identify geographic areas that have changed, the extent and distribution of the change, and to identify the nature of the change (e.g., “from-to” change information). The selection of an appropriate change PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Augu s t 2009 951 Jungho Im is with the Department of Environmental Resources and Forest Engineering, State University of New York, College of Environmental Science and Forestry, 1 Forestry Drive, Syracuse, NY 13210 ([email protected]). Jinyoung Rhee and John R. Jensen are with the Department of Geography, University of South Carolina, 709 Bull Street, Columbia, SC 29208 ([email protected]; [email protected]). Photogrammetric Engineering & Remote Sensing Vol. 75, No. 8, August 2009, pp. 951–961. 0099-1112/09/7508–0951/$3.00/0 © 2009 American Society for Photogrammetry and Remote Sensing Enhancing Binar y Change Detection Performance Using A Moving Thr eshold Window (MTW) Appr oach Jungho Im, Jinyoung Rhee, and John R. Jensen detection algorithm is one of the most important decisions in change detection investigations. Numerous algorithms and methods have been developed for remote sensingbased change detection (e.g., Malila, 1980; Michalek et al., 1993; Lyon et al., 1998; Lunetta et al., 2002; Im and Jensen, 2005; Im et al., 2007; Im et al., 2008). Sometimes, only binary change information (i.e., change versus no-change) is required especially when quick overview products are needed as preliminary output. Binary change detection generally consists of two steps: (a) creating changeenhanced images, and (b) selecting the appropriate spectral thresholds to generate binary change masks. Careful selection of spectral thresholds that discriminate “change” pixels from “no-change” pixels in change-enhanced images is necessary. Traditionally, binary change detection has been performed using a single change-enhanced image and a threshold using a trial-and-error approach. Im et al. (2007) developed an automated calibration model which removed three major limitations of the existing binary change detection procedures, by reducing manual intervention, using multiple change-enhanced images together, and testing continuous thresholds. The selection of appropriate change-enhanced images is also important in binary change detection. According to Im et al. (2007), change-enhanced images can be roughly divided into three groups: (a) a linear group, (b) a difference group, and (c) a ratio group. A linear change-enhanced image has the domain where one end is highly related to “change” and the other end is associated with “no-change” (e.g., change vector magnitude, neighborhood correlation coefficient). Difference and ratio types of change-enhanced images assume symmetry: a difference type (e.g., band difference) is assumed to be symmetric around the line of y 0 (i.e., center for symmetry 0), and a ratio type (e.g., band ratio) is assumed to be symmetric around a line of y x (i.e., center for symmetry 1). Many binary change detection studies using difference and/or ratio types of change-enhanced images have been based on such symmetry (e.g., Fung and LeDrew, 1988; Coppin and Bauer, 1996; Mas, 1999; Morisette and Khorram, 2000; Im et al., 2007). The assumption of symmetry, however, may not be reasonable in many real-world applications of binary 951-961_CD-5.qxd 7/20/09 4:13 PM Page 951