Change detection, accuracy, and bias in a sequential analysis of Landsat imagery in the Pearl River Delta, China: econometric techniques

Time series data from high resolution satellite imagery provide researchers with an opportunity to develop sophisticated statistical models of land-cover change. As inputs to statistical models, land-cover change data that are generated from satellite imagery must be both accurate and unbiased. This paper describes a new change detection method to determine the date of land-cover change in a sequential series of Landsat TM images of the Pearl River Delta, China. The method is a three-step change detection procedure that uses time series and panel econometric techniques. In the first step, regression equations are estimated for each of the six DN bands for each of seven stable land-cover classes. In the second step, the regression equations for each class are used to calculate DN values for change land-cover classes for each of the eight possible dates of change (1989-1996). In the third step, the date of land-cover change is identified by comparing a pixel's DN values against the eight possible dates of change using tests for predictive accuracy. The accuracy and bias of the dates of change identified by the econometric technique compare favorably to a more conventional change detection technique. Furthermore, the econometric technique may reduce efforts required to assemble the training data and to correct the images for atmospheric effects. Together, these results indicate that is possible to generate land-use change estimates from a time series of satellite images that can be used in conjunction with socioeconomic data to estimate statistical models of land-use change. (C) 2001 Elsevier Science B,V. All rights reserved.