Per-instance automated algorithm selection (AAS) aims at leveraging the complementarity of optimization algorithms with respect to different problem types. State-of-the-art AAS methods for numerical black-box optimization rely on supervised learning techniques that are supported by exploratory landscape analysis (ELA) feature sets. Recent works question the generalization ability of popular AAS approaches, which motivated the design of alternative feature sets. In this work, we take a closer look at the recently proposed set of Deep ELA features and investigate the ways in which Deep ELA complements the classical ELA feature sets. To this end, we first study the correlation between the two feature collections, both through pairwise classification and through regression models. The complementarity observed in these analyses is confirmed by an AAS study, where models combining deep and classical features outperform those that are restricted to selecting from only of the two collections.