Date of Graduation

2004

Document Type

Dissertation/Thesis

Abstract

Spatial structure in remotely sensed imagery is shown to impact the change detection procedure. Methods to incorporate spatial statistics into correlation-based and regression-based change detection that control for the spatial structure in this data are proposed. These procedures are presented by means of an urban change detection case study utilizing Landsat TM and Landsat ETM+ imagery in the area to the west of the Baltimore, MD I Washington D.C. metropolitan area between 1989 and 2002. In the correlation-based analysis, the results show that the use of point-to-point correlation is often inappropriate in the examination of spatial datasets. In addition, spatial datasets require techniques specifically designed to account for the unique properties of this type of data so that relationships between variables are not biased. Matrix comparison is one technique that explicitly incorporates the spatial nature of data in the analysis of such datasets. Furthermore, the matrix comparison approach also provides additional directional information about the correlation that can be used to explore and evaluate the relative merit of alternative hypotheses of underlying spatial processes affecting the analysis at hand. Traditional approaches to image regression change detection will produce unsatisfactory results by assuming the relationships in question behave in a consistent manner from place to place (i.e. the relationships are spatially stationary.) Geographically weighted regression (GWR) addresses this weakness by generating local parameter estimates for each observation in a particular study area. As this research shows, local GWR models typically provide higher coefficients of determination than their aspatial counterparts. In addition, the spatial variation commonly exhibited by the local GWR parameter estimates in this research reveals that traditional aspatial regression approaches not only fail to incorporate the spatial structure of the data into the analysis, but also the failure to do so may generate misleading results with respect to the regression residuals on which image regression change detection is based.

Download

Share

COinS