Deep learning for Computer Vision (CV) tasks such as object detection, classification, and segmentation have been tremendously successful in the decade since the introduction of Alexnet and GoogLeNet architectures. This success is attributable to large image training sets and increasingly complex network architectures which capture image features into representations in correspondingly finer detail. Image training sets in the handheld imagery domain typically have sufficient positional (or optical) separability between features within an image frame to make detectors effective. Satellite imagery, on the other hand, are taken from a further standoff distance where optical separability between image features are significantly smaller. Attempts to bridge the model performance gap between the two modalities have been pursued over the years, resulting in the utility of model architectures such as PyramidNet, Detectron2, and ResNet. However, the differences in learned feature representations within the trained model for both modalities have never been studied. Pose variances and short collection distance from a camera to subject in the handheld domain work well with existing computer vision methods, but these parameters are remarkably different in satellite imagery, thereby pushing the limits of the best vision models today. This study explores the limitations of encoded features from popular commercial satellite imagery training datasets such as those from SpaceNet (Maxar). Categorical features emphasized in these datasets (i.e. - aircraft parts such as wings, tails, engines, etc.) by their respective authors will serve as the basis for determining the limitations of minimum viable spatial feature separation for within hidden layers in neural networks. Finally, we examine the composability of features, and their separability in model representations consistent with Zernicke moments in computer vision research.