Combined spectral and spatial clustering with agents

The traditional approach of classifying multispectral and hyperspectral imagery begins with clustering in feature space, followed by labeling the classes in the image space. To overcome some of the disadvantages of this sequential approach is to consider spatial constraints during clustering, for example by using Markov Random Fields. We propose in this paper an agent based clustering method as it appears to more versatile for the purpose of identifying urban objects, such as buildings and roads. The proposed clustering method is based on the agent model developed during the last decade in artificial intelligence. Our approach utilizes not only the spectral information but also the spatial information like shape descriptors and distance descriptors for clustering. The method uses two agents, one for spatial clustering and one for spectral clustering. Both agents attempt to cluster the data first. Whenever one agent gets a new piece of information, the information is shared with the other agent and used to update the existing information. Through the communication mechanism, the spectral information and the spatial information are tightly coupled during the clustering processes. We tested the proposed method with a multi-spectral image of an urban scene. The band characteristics are similar to those of an IKONOS multi-spectral image. We also compare our method with a fuzzy based clustering method that uses only the spectral information, and elaborate on the performance improvement. The experimental results show that combining the spectral information with the spatial information increases the clustering performance.