Remotely Sensed Change Detection Based on Artificial Neural Networks

A new method for remotely sensed change detection based on artificial neural networks is presented. The algorithm for an automated land-cover change-detection system was developed and implemented based on the current neural network techniques for multispectral image classification. The suitability of application of neural networks in change detection and its related network design considerations unique to change detection were first investigated. A neural-network-based change-detection system using the backpropagation training algorithm was then developed. The trained four-layered neural network was able to provide complete categorical information about the nature of changes and detect land-cover changes with an overall accuracy of 95.6 percent for a fourclass (i.e., 16 change classes) classification scheme. Using the same training data, a maximum-likelihood supervised classification produced an accuracy of 86.5 percent. The experimental results using multitemporal Landsat Thematic Mapper imagery of Wilmington, North Carolina are provided. Findings of this study demonstrated the potential and advantages of using neural network in multitemporal change analysis. Introduction Global environmental change has become a major national and international policy issue. Not only does change alter the local landscape, but it may also produce ecosystem effects at some distance from the source (Dai and Khorram, 1998a). While a considerable amount of data about the nature of the Earth's surface has been collected by remote sensing devices, the volume and rate of these data are expected to increase rapidly as more images of various resolutions become available in the public domain, such as Earth Observing System [EOS) data (Asrar and Greenstone, 1995). These remotely sensed data are used to determine land use and land cover at a given point in time and land-cover changes between multiple dates (Miller et al., 1995). Given the current techniques available, remote sensing provides one of the most feasible approaches to local, regional, and global land-cover change detection (Khorram et al., 1999). Many change-detection techniques are used in practice today. Most techniques are semi-automated because analysts still have to manually carry out many image processing tasks such as image registration, threshold tuning, and change delineation. There are also problems associated with semi-automated techniques, including being time-consuming, inconsistent, and difficult to apply to large-scale and global information systems, such as the International Earth Observing System (IEOS) (Dai and Khorram, 1998b). Additionally, a number of the techniques can only provide a binary change mask, and a classification procedure must be applied to the multitemporal images to extract categorical change information (Serpico and Center for Earth Observation, North Carolina State University, Raleigh, North Carolina 27695-7106 ([email protected]). PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING Bruzzone, 1997; Coppin and Bauer, 1996; Singh, 1989). Therefore, a reliable automated change-detection system identifying categorical changes would be useful in environmental remote sensing and its regional or global implementation. This paper reports the development of procedures for such a change detection system based on artificial neural networks. This paper includes five sections. An overview of remotely sensed change detection is first presented. Experimental design of the proposed neural-network-based change-detection system is then discussed, which includes the network input, output, and architecture, along with fundamentals of the backpropagation training procedure. The experimental results are then presented, where we focus on the classification scheme, training data development, network parameter selection, generalization problems, change detection accuracy assessment, and comparison with other techniques of categorical change detection. Finally, conclusions and recommendations are given. Remotely Sensed Change Detection Usually, change detection involves two or more registered remotely sensed images acquired for the same ground area at different times. During the last two decades, there have been many new developments in remotely sensed change detection. These techniques may be characterized by their functionalities and the data transformation procedures involved. Based on these characteristics, we can classify current change-detection techniques into two broad categories: Change Mask Development (CMD): Only changes and nonchanges are detected and no categorical change information can be directly provided; and Categorical Change Extraction (CCE): Complete categorical changes are extracted. In the first categow, changed and non-changed areas are separated by a t6reshoB when comparing the spectral reflectance values of multitemnoral satellite images. The amount of change is a functionbf the preset thresGold. The threshold has to be determined by experiments. The nature of the changes is unknown directly from these techniques and needs to be identified by other pattern-recognition techniques. Therefore, these techniques are only suitable for development of a change mask. Most change-detection methods fall into the first category. For example, Lmage Differencing, Image Ratioing, and Image Regression only lead to the development of a change mask. These techniques can be used for data of one band, two bands, three bands, or more than three bands, with decision Photogrammetric Engineering & Remote Sensing, Vol. 65, No. 10, October 1999, pp. 1187-1194. 0099-1112/99/6510-1187$3.00/0 O 1999 American Society for Photogrammetry and Remote Sensing