SPIE Proceedings Vol. 4054, Aerosense 2000, Orlando, FL (in press) Automated spatiotemporal change detection in digital aerial imagery P. Agouris 1 , G. Mountrakis 1 , A. Stefanidis 2 1 Department of Spatial Information Engineering, University of Maine 2 National Center for Geographic Information and Analysis (NCGIA), University of Maine ABSTRACT Handling change within integrated geospatial environments is a challenge of dual nature. It comprises automatic change detection, and the fundamental issue of modeling/representing change. In this paper we present a novel approach for automated change detection which allows us to handle change more efficiently than commonly available approaches. More specifically, we focus on the detection of building boundary changes within a spatiotemporal GIS environment. We have developed a novel approach, as an extension of least-squares based matching. Previous spatial states of an object are compared to its current representation in a digital image, and decisions are automatically made as to whether or not change at the outline has occurred. Older object information is used to produce templates for comparison with the representation of the same object in a newer image. Semantic information extracted through an analysis of template edge geometry, and estimates of accuracy are used to enhance our method. This template matching approach allows us to integrate in a single operation object extraction from digital imagery with change detection. By decomposing a complete outline into smaller elements and applying template matching along these locations we are able to detect precisely even small changes in building outlines. In this paper we present an overview of our approach, theoretical models, certain implementation issues like template selection and weight coefficient assignment, and experimental results. Keywords: Change Detection, Object Extraction, Least Squares Matching, Spatiotemporal 1. INTRODUCTION Object extraction is a fundamental photogrammetric process. The resulting information, namely precise positional data for objects (e.g. buildings, roads) on the Earth’s surface, is essential to a large variety of applications, especially GIS-related ones. In traditional digital photogrammetric and computer vision approaches, object extraction from digital images comprises two major operations: identifying the object within an image, and precisely tracking this object by measuring the image coordinates of its outline. The first process involves a variety of logical decisions like image interpretation, understanding, and object classification. The latter is a precise localization problem, aiming at subpixel accuracies in the outline detection. Research in digital photogrammetry and computer vision during the last 20 years showed that there exist no universal edge detectors which can be applied to a digital image function to both identify and track edges with sufficient success. Instead, there exists a trade-off between: • reliability, which expresses the qualitative accuracy associated with identification, and • precision, which expresses the geometric accuracy associated with tracking. Operators that excel in reliability are often referred to as type I operators, in analogy to type I errors in statistics. They succeed in identifying classes of objects (e.g. houses) within an image, without particularly dealing with precise outline localization. Their products may be for example object blobs. Operators that excel in accurately positioning object outlines are often termed type II operators. Typically, they function successfully within narrow search windows. Operators from these two broad classes have often been combined, in an effort to optimize both accuracy measures 12 . A good overview of the ______________________________ Correspondence: Email: {peggy, giorgos, tony }@spatial.maine.edu , Telephone : (207) 581-2180 Address: 5711 Boardman Hall #348, Orono, ME, 04469, USA