The military use of the Internet of Things within a battlefield environment aims to combine the information collected from a system of heterogeneous sensors and actuators in order to create a cohesive model of the relevant battlefield so that intelligent agents can provide risk-aware decisions or take proper actions, and to collectively give warfighters an edge. To do inference and reasoning under uncertainty efficiently, the most important and relevant features regardless of modality must be identified for each given context, classification task, and classification approach. This can minimize the computational costs required to build a specific model, increase the accuracy of the model and may also allow the model to be generalized. However, the dynamic and adversarial nature of the battlefield may mean that the availability and reliability of sensors will vary over time. Adding a certain amount of redundancy in the set of features used to train an ensemble of classifiers may improve model robustness and minimize uncertainty. One approach to achieve this is by modeling the feature space so that the likely importance of a given set of features can be estimated when context, classification task, or approach is varied. To efficiently understand the shape of a given feature space and to locate clusters of features in a locally distributed fashion, we surveyed methods to select important features and to describe or explore a given feature space.