4th International Conference on Integrating GIS and Environmental Modeling (GIS/EM4):
Problems, Prospects and Research Needs. Banff, Alberta, Canada, September 2 - 8, 2000.

Change Detection for Urban Growth Modeling

an artificial neural network approach

GIS/EM4

XiaoHang Liu

Abstract

A new method based on artificial neural network (ANN) is presented to detect the change from non-urban to urban land use. Using Levenburg-Marquart algorithm, the three-layered backpropagation network was able to identify the changes of interest with an overall accuracy of 93.10%. Compared to conventional change detection algorithms such as post classification comparison, ANN is 40% more accurate when Kappa coefficients was employed. The Levenburg-Marquart algorithm enables ANN to converge in less than 11 minutes on a Pentium 266 CPU. The intensive training cost which often prevents the practical application of ANN was thus not a constraint.

Keywords

Urban growth modeling, change detection, artificial neural network, Levenburg-Marquart algorithm, post classification comparison

Introduction

Population growth, together with technological advancement, has changed many natural ecosystems into human development. To understand the past as well as to plan for the future, many urban growth models have been constructed. In calibrating and evaluating these models, satellite imagery is often employed to obtain land use change information. Considering the impact of misidentified changes on modeling accuracy, an accurate change detection algorithm is necessary. Moreover, the complexity of the algorithm should be as minimal as possible as usually a sequence of change detection has to be performed.

The primary type of land use change interesting to the urban modeling community is from non-urban to urban land use change. Existing algorithms such as post classification comparison and image differencing are not designed specifically for this task. As a result, either the accuracy is not satisfying or efforts involved are intense. For example, when post classification comparison is applied, each single-date imagery has to be classified into detailed classes before they are aggregated up and compared to obtain the change information. Omission-commission errors resulted from single date classification is often significant (Jensen 1996).

This study explores the utility of artificial neural network to detect the change from non-urban land use to urban land use. Traditionally, ANN has been only applied to perform single date imagery classification. Among the few explorations to apply ANN to change detection(Gopal and Woodcock 1996, Dai and Khorram 1999) , few have addressed ANN’s intensive training cost which forms the major constraint for its practical application. This study addresses this issue. For comparison purposes, post classification comparison was also performed.

Methods

Two Landsat TM imagery of Barnegat Bay, New Jersey, were acquired for 1984 and 1994. Both images were georectified and registered to UTM coordinate system. During post classification comparison, each single date imagery was classified into 7 classes: developed area, cultivate/grass land, woodland, bare land, Marine/Esturine emergent land, River/Lacustrine/Paulstrine wetland and open water. The latter 6 classes form the non-developed area. By cross-tabulating the two classified imagery, a “from-to” change matrix was obtained. The training and test pixels used for post classification comparison are identifical to those for artificial neural network.

Crude stacking of original TM images will result a composite image of 14 bands. High bands require increasing number of training data to ensure the classification reliability. To avoid this effect, a standardized principal component analysis (PCA) was applied to each single date imagery. In both cases, the first three bands were found to account for 97% of the total information. The resulted composite imagery is of 6 bands then.

To classify the composite imagery, a feed-forward network with a backpropagation learning rule was applied. The ANN consists of 3 layers, with 6, 10 and 1 node respectively. A tangent sigmoid function is used as the activity function. Levenburg-Marquart algorithm, one of the fastest backpropagation methods were applied to train the network. During the training state, training pixels were presented to the network one after another. The network compares its prediction with the target to adjust its estimation iteratively. For comparison purposes, post classification comparison was also performed with maximum likelihood as the classifier.

Findings & Conclusion

Compared to post classification comparison, ANN was 40% more accuate when Kappa coefficient was used to measure change detection accuracy. In all trials, Levenburg-Marquart algorithm enables ANN to converge in less than 1200 iterations. In a personal computer with Pentium 266 CPU, this is less than 11 minutes. The intensive training cost which often prevents the practical application of ANN was thus not a constraint. Results are listed in the following table.

References used

Jensen J. 1996. Introductory digital image processing: a remote sensing perspective, 2nd edn, New Jersey: Prentice-Hall.

Dai, X.L., and Khorram, S., 1999. Remotely sensed change detection based on artificial neural networks. Photogrammetric Engineering & Remote Sensing, 65:1187-1194.

Gopal, S. and Woodcock, C., 1996. Remote sensing of forest change using artificial neural networks. IEEE Transaction on Geoscience and Remote Sensing, 34:398-403.

Authors

XiaoHang Liu, Ph.D. student, Department of Geography
University of California at Santa Barbara
Email:xhliu@geog.ucsb.edu, Tel: +1-805-893-8652, Fax: +1-805-893-8617.