A Matching Algorithm for Detecting Land Use Changes Using Case-Based Reasoning

The paper deals with change detection using time series SAR images. SAR provides a unique opportunity for detecting land-use changes within short intervals (e.g., monthly) in tropical and sub-tropical regions with cloud cover. Traditional change detection methods mainly rely on per-pixel spectral information but ignore per-object structural information. In this study, a new method is presented that integrates object-oriented analysis with case-based reasoning (CBR) for change detection. Object-oriented analysis is carried out to retrieve a variety of features, such as tone, shape, texture, area, and context. An incremental segmentation technique is proposed for deriving change objects from multi-temporal Radarsat images. Feature selection based on genetic algorithms is carried out to determine the optimal set of features for change detection. A CBR matching algorithm is developed to identify the temporal positions and the kind of changes. It is based on the weighted k-Nearest Neighbor classification using an accumulative similarity measure. The comparison of the four combinations of change detection methods, object-based or pixel-based plus case-based or rule-based, is carried out to validate the performance of this proposed method. The analysis shows that this integrated approach has provided an efficient way of detecting land-use changes at monthly intervals by using multi-temporal SAR images. Introduction Since the launch of ERS-1 (1991) and ERS-2 (1995) by the European Space Agency (ESA) and Radarsat-1 (1995) by the Canadian Space Agency (CSA) and NASA, the synthetic aperture radar (SAR) has become a weather-independent monitoring tool covering a large part of the globe on a regular basis. There is an increasing demand for using satellite SAR images as a complementary data source for resource inventory and environmental monitoring (Ranson and Sun, 2000; Baghdadi et al., 2001; Magagi et al., 2002). This is because conventional optical remote sensing may PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Novembe r 2009 1319 Xia Li and Bin Ai are with the School of Geography and Planning, Sun Yat-sen University, 135 West Xingang Rd., Guangzhou 510275, P.R. China ([email protected]; [email protected]). Anthony Gar-On Yeh and Zhixin Qi are with the Centre of Urban Planning and Environmental Management, The University of Hong Kong, Pokfulam Road, Hong Kong SAR, P.R. China. Jun-ping Qian is with the School of Geography and Planning, Sun Yat-sen University, 135 West Xingang Rd., Guangzhou 510275, P.R. China, and the Guangzhou Institute of Geography, Guangzhou 510070, PR China. Photogrammetric Engineering & Remote Sensing Vol. 75, No. 11, November 2009, pp. 1319–1332. 0099-1112/09/7511–1319/$3.00/0 © 2009 American Society for Photogrammetry and Remote Sensing A Matching Algorithm for Detecting Land Use Changes Using Case-Based Reasoning Xia Li, Anthony Gar-On Yeh, Jun-ping Qian, Bin Ai, and Zhixin Qi have difficulties in collecting the desired ground data in regions frequently affected by clouds; hence, it cannot be used for monitoring land-use changes, especially for identifying illegal land development in fast-growing regions in tropical and sub-tropical areas that often experience cloud cover. Monitoring land-use changes in these areas on a regular basis (e.g., monthly) can help to prevent illegal developments at an early stage. Monthly SAR images can obtain ground information almost in real time and thereby provide an efficient tool for monitoring land-use changes and urban development in rapidly developing regions. Monitoring land-use changes at short intervals can be based on the independent classification of these monthly SAR images. Post-classification comparison can be used to identify not only the amount and location of change but also the nature of change (Howarth and Wickware, 1981; Richards and Jia, 1999). However, such comparison has limitations because the classifications of individual images always contain errors and the resulting change-detection analysis can end up being more a classification of error than a classification of change. As a result, the degree of land-use changes may be overestimated by comparing a number of independent classifications (Li and Yeh, 1998). Moreover, this type of method faces difficulties when a long sequence of time series data is involved. A common method is to apply principal components analysis (PCA) to obtain information on land-use changes from a long sequence of time series data (Eastman and Fulk, 1993; Li and Yeh, 1998). Although many studies exist on the methodologies of change detection, only a few have been devoted to change detection using SAR images, mainly because of the intrinsic complexity of SAR data. Recently, some studies have been published on the development of algorithms using SAR images for change detection (Bovolo and Bruzzone, 2005). However, there is still a general lack of studies that focus on the use of object-oriented analysis to retrieve spatial information from SAR images for change detection. Orbital radar images are at present often obtained using only one single frequency (e.g., C-band). Significant confusion arises if land-use classification and change detection are based purely on the information of a single band of these SAR data (Li and Yeh, 2004). The first way to reduce such confusion is to use time series SAR images, since temporal information can partially compensate for the limitations of using a single frequency. The second way is to derive ancillary features related to structural information,