An unsupervised marker image generation method for watershed segmentation of multispectral imagery

Watershed transformation in mathematical morphology is a powerful tool for image segmentation. Watershed transformation based segmentation is generally a marker-controlled segmentation. This paper proposes a novel method of maker image generation based on unsupervised ISODATA classification. The method incorporates spectral information contained in multispectral data into watershed transformation for image segmentation. The method is evaluated and compared to existing unsupervised method. The results show that the proposed method could effectively reduce the oversegmentation effect and achieve more accurate segmentation results, compared to existing method. The combination of segmentation result and pixed-based classification is found to improve the overall classification accuracy over pixel-based classification.