Very high spatial resolution (VHSR) commercial satellite imagery affords permafrost scientists the ability to monitor the pan-Arctic system at a fine-scale, enabling detailed monitoring of both the natural and human environments. Geo-AI mapping applications based on the deep learning (DL) convolutional neural network (CNN) have been successful in translating this big imagery resource into Arctic science-ready products. However, many models are computationally intensive due to the constraints of the large geographical extent and complexity of VHSR imagery. In addition, feature recognition is challenged by scarcity of manually-annotated training data and image complexity at fine scales. In this exploratory study, we investigated the ability of a lightweight U-Net DLCNN to efficiently perform semantic segmentation of VHSR commercial satellite imagery with limited training data in automated recognition of human-built infrastructure, including residential, industrial, public, commercial buildings, and roads, in the permafrost affected regions of the Arctic. We conducted a systematic experiment to understand how image augmentation improves the performance of DL-based semantic segmentation of VHSR imagery. Different standard augmentations, including flipping, rotation, and transposition, were applied to input imagery in order to test their impacts on infrastructure recognition and determine the optimal set of augmentations. With a relatively low number of model parameters, limited labelled training data, short training time, and high segmentation accuracy, our findings suggest that overall, the U-Net DLCNN, coupled with image augmentation, could serve as an accurate and efficient method for mapping infrastructure in the Arctic permafrost environment without compromising spatial details and geographical extent.