Measured under-ice acoustic profile data and several large-scale geometric parameters of ice keels are used to construct and partition an ensemble of large-scale relief features into two subsets: a subset of keel-like features (e.g., ice keels) and a subset of nonkeel-like features. The draft data of each feature are regarded as a realization of a nonstationary random process while the draft increment data are regarded as a realization of a stationary zero mean random process. A maximum likelihood estimator technique (MLE) and a technique based on the variance function are used to calculate fractal dimensions for keel-like and nonkeel-like features. It is shown that for the same feature, the MLE technique and the variance function based technique yield similar values for the fractal dimension (D) and that for keel-like features 1.2