Clustering is like any segmentation method, but it is an unsupervised method which requires no pre-training to recognize information in the image. The correction of the under or over segmentation can be controlled using a combination of clustering with another method. So, I guess using clustering to improve watershed segmentation comes from this objective because clustering provides a good initialization for the algorithm to start correctly.

The performance of marker controlled watershed algorithm can be improved by putting appropriate filtering algorithm prior to segmentation. Also when using clustering along with watershed algorithm, for cluster centroid initialization, optimization algorithm can be employed for obtaining better results.