The unique, bi-pedal motion of humans has been shown to generate a characteristic micro-Doppler signature in the time-frequency domain that can be used to discriminate humans from not just other targets, but also between different activities, such as walking and running. In the literature, many different features have been proposed for classification applications. However, it is not known which features have a greater impact on classification performance, or indeed how many features should be used to achieve good classification. In this work, the mutual information of features extracted from human micro-Doppler signatures is computed. Taking the problem of classifying human arm-swing as an example, the features extracted are ordered in terms of importance.