Identifying Objects

Abstract

Object identification is often a key step in data mining spatial grids. Information may need to be extracted not for individual pixels but for groups of pixels that together form a real-world entity such as a mountain range, city, or a thunderstorm. Commonly, objects are formed by combining high-valued pixels that are connected into objects. This is done using a process called region growing. Once candidate objects have been identified, it is possible to compute geometric and physical attributes of the regions to fine-tune the object identification. Objects can also be geocoded based on their centroids. An improvement to using a single, hard threshold is to use hysteresis to provide a softer threshold. Another way to provide softer criteria to define a region is to employ active contours or snake algorithms. However, all these techniques rely on global thresholds. The watershed transform allows for local, adaptive thresholds using an immersion metaphor. An enhancement of the watershed transform, suitable for many real-world spatial grids, is to specify objects in terms of a minimum size requirement. The final object identification technique presented in this chapter is contiguity-enhanced clustering which has the advantages of using local, adaptive thresholds while not relying on pixel connectivity.