ABSTRACT Incorporating satellite imagery is crucial for remote sensing and computer vision in socio-economic studies. Machine learning techniques are typically used to extract valuable information from satellite images. This article presents an enhanced approach applying computer vision not only to satellite images but also to five different map sources, such as OpenStreetMap and building footprints. The goal is to determine if useful insights can be derived from simplified feature representations, improving the understanding of fundamental satellite imagery data. We conducted an experiment predicting the settlement patterns of university graduates in Vienna, using a convolutional neural network (CNN) to analyze grid cell images (250 m × 250 m) from satellites and five different maps. The model predicted five density classes of graduates, achieving an accuracy rate of 35.99% using building footprints, outperforming the 35.15% accuracy based on satellite images, while other map representations underperformed. These results suggest that building outlines and the open space between buildings contain vital predictive information. Our findings highlight the potential of this approach beyond socio-economic variables, demonstrating the capability of understanding maps via CNNs.