Existing nationwide spatial datasets do not adequately represent Germany’s traffic infrastructure. They are often fragmented, vary in quality, and lack documentation on acquisition methods and spatial coverage. This is especially true for spatial data on parking areas, despite the growing demand for such information in traffic management and urban planning. The topic is gaining importance in the context of repurposing on- and off-street parking in cities (Thigpen and Volker, 2017). While some large cities like Berlin maintain spatial parking inventories, most smaller cities do not. In Berlin, data is collected via vehicle-mounted camera systems (Senatsverwaltung für Mobilität, Verkehr, Klimaschutz und Umwelt, 2023). However, this approach only maps public parking areas, is costly in time and resources, and excludes private and semi-private spaces. Large-scale mapping using aerial imagery can help to bridge this gap, enabling efficient coverage from city to national scale. This study presents a novel parking area inventory for Berlin derived from aerial imagery by AI-based image analysis methods. Results include public, semi-private, and private parking areas, enriched with information on accessibility and capacity.

The methodology is based on aerial imagery with a ground sampling distance of 10 cm, such as those provided for example by the German state surveying offices (DOP10). Traffic areas are segmented using a widely-successful and robust segmentation model based on the U-Net and DenseNet architectures (Henry et al., 2021). It leverages the ability from the former to extract fine-grained information from every feature level of input images, while benefitting from the latter’s efficient information flow and optimization in both the encoder and decoder. It was trained and validated on the novel TIAS (Traffic Infrastructure and Surroundings) dataset. TIAS consists of 51 manually annotated images from diverse traffic environments across German cities – 45 acquired using DLR’s 3K and 4K camera systems, and six DOP10 images from public authorities. It includes nine classes: road, access way, bikeway, footway, keep-out area, parking area, railroad bed, road shoulder, and water (Merkle et al., 2024). For this study, only the classes parking area, road, and access way were segmented.

The resulting segmentation maps are vectorized, smoothed, and imported into a PostGIS database. Each parking polygon is assigned an access category by intersecting with cadastral data on land parcel usage (Senatsverwaltung für Stadtentwicklung, Bauen und Wohnen Berlin, 2025). These usage types are grouped into public, semi-private, and private access levels and applied to the respective polygons.

To estimate the number and positions of parking spots from each vectorized parking polygon, the type of parking – parallel, diagonal, or vertical – must be determined, as each type requires different dimensions according to German parking layout standards (FGSV, 2023). The estimation involves three steps: (1) extracting centerlines of roads and accessways, (2) identifying vehicles and their orientation relative to the road, and (3) assigning a parking type based on average orientation of vehicles on a parking polygon.

Centerlines of each road and accessway polygon are extracted using PostGIS’s ST_VoronoiLines function, followed by a three-stage filtering process to remove support lines towards the polygon edges. This yields one centerline per polygon. Light- and heavy-duty vehicles are detected using a transformer-based object detection framework – DINO, as adapted by Mühlhaus et al. (2023) for oriented bounding box detection in the aerial imagery domain, replacing the standard horizontal bounding box representation. The model was trained on the Eagle dataset – an aerial dataset for vehicle detection (Azimi et al., 2020). The detected objects are subsequently stored in a PostGIS database for further analysis. The longitudinal axis of each vehicle is extended five meters front and back, then intersected with centerlines of roads and accessways to obtain the intersection angle. For vehicles without direct intersections, the mean deviation between their orientation and nearby centerline segments within a buffer is used. This is especially relevant for parallel parking.

The mean orientation of all vehicles on a parking polygon defines its dominant orientation. Based on this angle, parking types are classified as: parallel (0°–30°), diagonal (31°–75°), and vertical (76°–90°). Parking capacity is calculated by dividing the area of the polygon by standard slot dimensions from FGSV (2023): parallel (5.5 m × 2.15 m), diagonal (5.42 m × 2.7 m), and vertical (5.14 m × 2.7 m). For polygons without identified parking type, a default slot size of 5.2 m × 2.4 m is used.

The resulting Berlin-wide inventory comprises 1,333,953 parking spots. Of these, 60% are classified as publicly accessible, 21% semi-private, and 19% private. This reveals that inventories limited to public parking underestimate total availability by around 40%. Regarding layout, 36% of the spots are parallel, 27% diagonal, and 20% vertical; 17% could not be classified due to lack of vehicle data.

This study demonstrates the potential of aerial imagery for generating comprehensive parking inventories. The method is efficient, highly automated, and scalable, making it applicable to larger areas. Moreover, the results highlight the limitations of current approaches that only capture public parking. Future work will address larger-scale processing, integration of underground parking data, and development of correction factors for partially occluded areas.