For the watershed segmentation algorithm to be successful it must be implemented on a realistic gradient image. In most watershed implementations, gradients are extracted using an operator optimal for ideal step edges. However, image edges are never ideal steps and more closely resemble ramp edges at multiple scales. Therefore this strategy results in an inaccurate measure of image gradients and in turn lessens segmentation performance. In this paper we propose a new multiscale gradient operator for ramp edges. This strategy merges the properties of accurate gradient estimation of a large scale kernel with accurate localization of a small scale kernel by tracking gradients from larger to smaller scales. Quantitative performance evaluation of segmentation results shows this approach to outperform a traditional single small scale gradient operator optimal for step edges.