Detecting Residential Regions by Graph-Theoretical Measures

Abstract

We next consider a more specific land class type, namely “detecting residential regions”, in this chapter of the book. Unfortunately, spatial coherence based method introduced in the previous chapter is not sufficient for such fine classification. Our graph-theoretical measures, explained in Chap. 8 in full detail, can be used directly to solve this hard problem. We focused on sparse residential areas here. These areas, often including large, overhanging trees, exhibit forest-like characteristics and present a significant challenge. Yet, they are also of particular importance because it is often in these areas that growth most rapidly occurs. To detect these regions, we implemented and tested two strategies. In the first, we developed a Neyman–Pearson decision system based on the Mahalanobis distance to the center of the residential region’s distribution in our graph-theoretical feature space. The second approach is to define a three-class problem (rural, residential, urban) with a Bayes classifier. We obtained very good results in both cases. It is particularly noteworthy that the residential regions show a low miss rate. We tested 295 images taken from different parts of the US, representing a wide range of regions, climates, and terrain. We will first consider the problem of detecting residential regions only (i.e., residential vs. nonresidential classification). Then we will recast the problem as a three-way classification into rural, residential, and urban.