Local linear statistical modeling as a framework for change detection Sandra E. Thompson, * Don S. Daly, Eileen M. Perry, Kevin K. Anderson Pacific Northwest National Laboratory ABSTRACT Within the context of remote sensing, change detection refers to the characterization of changes in a scene or on the earth’s surface with respect to a target or targets. Many fields, including agriculture, urban planning, and national security, use satellite imagery to detect changes. However, current approaches have two major limitations: 1) the methods are effective in identifying scene changes only if changes (i.e., angle, illumination, atmospheric conditions) due to viewing the scene are negligible or corrected for, and 2) the results are typically a binary response, i.e., change or no change. We introduce a method that applies linear statistical modeling locally to estimate scene changes while accounting for image viewing changes. In this method, we explicitly describe the quantitative variation at a pixel in terms of its pixel neighbors, spatial and temporal, with a linear probability model. We then estimate the model parameters with least squares, calculate modeling diagnostics for each neighborhood, and display these results as images in order to assess scene changes, viewing effects, and modeling effectiveness on a neighborhood scale. By focusing on image neighborhoods, this method can effectively separate scene changes from changes due to viewing the scene. Furthermore, the flexibility of the local linear model and the ease of calculation of parameter estimates and diagnostics allow the researcher to emphasize and efficiently quantify specific types of scene or viewing changes. Keywords: Change detection, linear statistical model 1. INTRODUCTION Within the context of remote sensing, change detection refers to the characterization of changes in the earth’s surface with respect to some target or targets of interest. Examples include identification of changes in vegetation, characterization of earthquake damage, identification of change in medical applications, tracking of targets, and measurement of urban growth. Detecting change in images has received a large amount of attention in the last 20 years. The techniques in change detection vary; they include methods that use the original images (differencing, ratioing), methods that use mathematical transformations of the original images (principal component analysis [PCA]), methods that create thematic maps and identify changes from the different thematic classifications through time, and combinations of these techniques. Most change detection techniques apply these methods either to the entire image with global parameters or to an image on a pixel- by-pixel basis, leaving the important neighboring pixel information untapped. The challenge in change detection is to identify changes of interest, among a host of noninteresting changes, such as identifying new urban growth in an agricultural area across seasons. Many of the noninteresting changes are due to the action of viewing the * sandy.thompson@pnl.gov; phone 1-509-375-6611; fax 1-509-375-2604; http://www.pnl.gov; Pacific Northwest National Laboratory, P.O. Box 999, MSIN K5-12, Richland, WA 99352.