Up-to-date road data is essential for everyday applications such as routing and navigation. The timeliness of road data is particularly crucial in disaster scenarios when the road network is simultaneously subject to numerous changes due to destruction, flooding, or blockages [1]. However, existing routing applications can only take these changed conditions into account after a significant delay. Official data is provided with long update cycles and is not consistent across national borders. Even commercial providers need time to process their data and offer updates. This is where OpenStreetMap (OSM) can play to its strengths [2]. Thanks to the involvement of the large OSM community, the OSM database can be updated quickly. Furthermore, due to its free availability (Open Data License), OSM data is already the first choice for many humanitarian organizations and companies for map products and traditional routing applications.

Nevertheless, existing applications cannot fully exploit the potential of OSM data in terms of its timeliness since detecting road changes solely based on OSM updates is sometimes inherently unreliable. This is because OSM is geographical data gathered primarily by volunteers, in which road completeness and geographic accuracy vary substantially across regions and mapping communities [3]. Also, not all changes in OSM automatically lead to an improved data foundation. While deliberate mis-mapping and vandalism are relatively rare phenomena in OSM, they have potentially serious consequences. In recent years, there has been a rise in OpenStreetMap vandalism often linked to geopolitical conflicts. Such inconsistencies motivate the exploration of cross-validation approaches, particularly leveraging computer vision (CV) models to extract road changes from satellite images or orthophotos, thereby enabling the cross-validation of results derived from OSM to enhance reliability.

While most academic research on OSM relies on snapshots that examine specific time points, studying it in real-time induces new research questions. In our case, we want to examine to what extent the stream of OSM changes can be matched temporally and spatially with changes observed in remote sensing imagery.

It is increasingly possible to rely on high-resolution satellite data to verify changes in OSM road data, particularly in emergencies [4,5]. However, processing multi-temporal satellite data can be a time-consuming task, and there is no easy way to determine, in near real time, which data in OSM has been modified and how the relevance of these changes to routing applications can be assessed. Consequently, we are analyzing how road changes can be immediately detected and assessed using OSM and RS data together.

We propose a monitoring system for the road network based on OSM and remote sensing (RS) data: OSM Road Monitor. The system processes all changes to the OSM road network and evaluates them for errors and their impact on routing results. To achieve this, OSM Road Monitor uses road change detection in zero-shot settings using satellite imagery and orthophotos. Thereby, the area of change is identified in RS data, and this area is contrasted with the changes made in OSM. In this paper, we present the concept towards a near-real-time road monitoring system and the first results of change-event detection for identifying changes between two aerial or satellite images taken at different times. By combining change events with OSM updates, this allows us to identify and verify changes in the road network.

Therefore, we conducted a qualitative evaluation of several pretrained CV models to investigate their zero-shot generalization capabilities. The first category (C1) refers to road changes by comparing road segments of each timestamp estimated by models pretrained for road segmentation [6]. C1 methods are attractive due to the high availability of pretrained models for road segmentation. However, C1 is intrinsically sensitive to compounded error. Any false positive or false negative at either timestamp can be misread as a change. The second category (C2) uses the same strategy as C1 but employs foundation models such as the Segment Anything Model (SAM) to leverage their zero-shot generalization capabilities [7]. Despite SAM’s superior performance across various datasets, its open-ended segmentation paradigm necessitates appropriate prompting techniques. Experimental evidence indicates that SAM variants with visual prompts are more suitable for road segmentation than those with pure text prompts [8]. Nevertheless, as C2 retains a similar inference pipeline to C1, it continues to suffer from compounded errors. Furthermore, the segmentation results are unreliable when changes in the OSM database are missing or incorrect. The third category (C3) directly adopts change detection by estimating change masks from two temporally different images, thereby effectively avoiding the compounded errors observed in C1 and C2. However, available pretrained models are typically optimized for building change detection rather than road network dynamics. Admittedly, changes to buildings may occasionally require updates to the surrounding road network, but this is not always the case. Furthermore, since the number of buildings significantly exceeds that of roads, the overall efficiency of this approach is theoretically low. Regardless of the methodological category, challenges inherent to the data itself should also be considered. This includes inter-temporal heterogeneity, resolution and acquisition variation, orthophoto occlusion effects such as shadow, canopy, and sheltering, and the scarcity of labeled road-change samples.

In this paper, we present the first results that are aimed at building a demonstrator of OSM Road Monitor to form a near-real-time monitoring system for the road network based on OSM and RS data. As proof of concept, we provide exemplary results with orthophotos from the Swiss Federal Office of Topography swisstopo [9]. A pretrained model that was placed first in the SpaceNet5 challenge is used for inference [6, 10]. It comprises an ensemble of eight models, each incorporating a U-Net-style decoder with either a ResNet50 or SE-ResNeXt50-32x4d backbone, containing 73.3M and 75.3M parameters per model, respectively. Figure 1 shows road masks and segmented orthophotos for 2022 and 2025, respectively, around 47.5530°N, 8.5350°E (Tile 2682-1267) in Switzerland. A newly constructed road (a) and demolished roads around the roundabout (b) are successfully detected. Hence, this paper argues that a multimodal, data-centric strategy is essential for effective road change detection. The demonstrator is publicly available on GitHub at https://github.com/tum-bgd/2026-SOTM-osmroadmonitor.

OSM Road Monitor is envisioned as a notification system. Users of the application are notified at defined intervals or as soon as relevant changes occur in the OSM road network. The changes about which users wish to be notified depend heavily on the specific use case, including activity, vandalism, road topology, and road condition. Accordingly, the demonstrator will be extended to a dashboard that allows a user to specify a region of interest and generate the required reports on demand. Thereby, the area of change is identified in the RS imagery and contrasted with the changes made in OSM. If the number of changes in both does not align, a flagging function raises a message indicating the discrepancy. Accordingly, we aim to develop a real-time monitoring system for the road network based on OSM and RS data.

This project is funded by Zentrales Innovationsprogramm Mittelstand (ZIM) of the Federal Ministry for Economic Affairs and Climate Action under grant number 16KN113530.