Swarm intelligence for object-based image analysis

Abstract

Object-based image analysis (OBIA) approaches are often superior to the pixel-based classifications at very high resolution (VHR) remotely sensed images. Due to the similar spectral signatures of land-cover classes especially in urban areas, spatial information must be exploited to produce improved classification maps at finer resolutions. Segmentation and rule-based classification are the main two steps of the powerful OBIA approach which is widespread in pattern recognition and classification applications. Selecting the best values for segmentation parameters has an important effect on the segmentation results. Once the image objects are derived, topological relations between them, statistical summaries of spectral and textural features and shape features can all be employed in the rule-based classification. Optimum feature selection also has an essential role for rule set generation. Thus, optimal parameter/feature selection may be an important process in both steps of the OBIA approach. Among other optimization techniques, metaheuristic optimization algorithms such as swarm-intelligence-based methods are very capable of solving feature selection problems. So, they can be used inboth steps of OBIA approaches. In segmentation, the capabilities of swarm intelligence may optimize the parameters. Moreover, ant colony optimization (ACO) and particle swarm optimization (PSO) are successfully utilized for optimum feature selection in forming rule-based classification. In the first two sections of this chapter, the necessity of performing OBIA in object recognition based on VHR images is explained. Then, the basis of optimum feature selection and optimization algorithms is mentioned. After a comprehensive review on the concepts of ACO, PSO and firefly algorithm (FA) as the powerful swarm intelligence algorithms, the capabilities of these algorithms are investigated in the field of optimum feature selection. Finally, the experimental results of performing the FA and PSO for optimum feature selection are investigated and generalized for improving the capabilities of the OBIA approaches.