Supervised Region-Based Segmentation of Quickbird Multispectral Imagery

The segmentation of very high resolution (VHR) satellite imagery (such as Digital Globe QuickBird) is becoming increasingly important to geo-related applications. New sensors provide the ability to discriminate large scale objects that were not discernable with lower resolution satellite imagery such as Landsat TM. VHR satellite images also exhibit an incredible dynamic grey-value variety. These features, among others, impede existing algorithms developed for lower resolution satellite imagery to operate within the same degree of accuracy. This paper proposes a supervised approach to the segmentation of QuickBird multispectral imagery through the integration of the Hierarchical Split Merge Refinement (HSMR) framework. The HSMR framework was originally developed by Ojala and Pietikainen [1999] for unsupervised segmentation of textured areas. In this approach, user identified regions are employed to guide HSMR algorithmic processes. User knowledge is brought to segmentation and it is hypothesized that this will improve stabilization in HSMR segmentation across a variety of QuickBird 2.44 m multispectral satellite image scenes and improve control of segmentation at different scales.