Transport-related surface detection with machine learning: Analyzing temporal trends in Madrid and Vienna

Abstract

This study investigates the integration of machine learning into urban aerial image analysis, focusing on identifying mobility-related surfaces for cars and pedestrians, such as parking spaces, road surfaces, and sidewalks, while also analyzing historical trends. It highlights the potential of foundational models and the advantages of fine-tuning them for global geospatial analysis. A workflow is proposed for automatically generating geospatial datasets, facilitating the creation of semantic segmentation datasets from diverse sources, including WMS/WMTS links, vector-based cartography, and OpenStreetMap (OSM) overpass-turbo requests. The developed code streamlines dataset generation for training machine learning models using publicly available data, eliminating the need for manual labeling. Two datasets for car- and pedestrian-related surface detection were generated using aerial imagery and vector data from the geographical offices of Madrid and Vienna, with a particular focus on parking surfaces. A transformer-based model was trained and evaluated for each city, achieving strong accuracy with F1 scores exceeding 0.5 in most classes in just less than 10 epochs. Historical trend analysis was conducted by applying the trained model to images from 10 to 20 years ago, prior to the availability of vector data, successfully uncovering temporal infrastructure trends for cars and pedestrians in different city areas. The trends discovered in Madrid and Vienna showed distinct differences: Madrid's periphery is experiencing a slower growth in parking surface, while Vienna is seeing a greater increase in parking compared to newly built road surfaces. This method offers a cost-effective solution for municipal governments to gather valuable urban data.