Geographical databases such as OpenStreetMap (OSM) are continuously evolving to reflect ground-level transformations. Identifying, describing, and characterizing these changes is a key challenge for analyzing territorial evolution over time. While several works have analyzed OSM edit history for quality assessment [1] or contributor behavior studies [2], the semantic interpretation of changes and, in particular, the ability to distinguish genuine ground-level evolution from data refinements, tagging corrections, or modeling decisions, remains largely unaddressed. For planners relying on OSM, distinguishing real infrastructure changes from editing artifacts is critical. We address this challenge through two contributions: a web application to explore change dynamics and a methodology for semantic annotation and interpretation of edits.

Our data source is the Planet file with history, which encodes every version of every object ever created in OSM since 2007 [3]. From this corpus, we use the ohsome-planet tool [4] to extract the complete edit histories of three study cities (Lisbon, Grenoble, and Paris) covering the full OSM timeline. The methodology is currently instantiated on cycling infrastructure, enabling the study of whether and how cities are progressively transitioning toward soft mobility, by tracking how cycling infrastructure is documented and refined in OSM over time. It is designed for generalization to any OSM object family.
We identified all tags related to cycling over the study period and selected a subset based on their frequency and their consistency with established OSM tagging practices. This curated set forms the basis for tracking tag usage and its evolution over time. These tags are then grouped into four semantic categories reflecting the main families of cycling infrastructure: Dedicated cycleways (e.g., highway=cycleway) Shared-use paths (e.g., highway=path + bicycle=designated), On-road lanes (e.g., cycleway=lane), and Designated routes (e.g., oneway:bicycle=no). Within this framework, we also distinguish two tiers of OSM tags: primary attributes, whose values determine which cycling infrastructure category an object belongs to, and secondary attributes, which capture additional properties of the object (e.g., street name, surface type, width). This distinction drives the interpretative pipeline: transitions at the main attribute level are considered significant events and are the primary focus of our analysis, as they capture the most meaningful shifts in how cycling infrastructure is represented and potentially built. Changes to secondary attributes provide complementary context for interpreting these dynamics.

The pipeline operates across two levels of analysis: (1) the observable level and (2) the interpretative level. At the observable level (1), changes between successive feature versions are conceptualized as transitions and classified along five dimensions: existence (creation, deletion), category, attribute (tag addition, modification, removal), geometry (transformation (e.g., node-to-way), geometry modification, densification, sparsification), and structure (fragmentation, aggregation, recomposition). This ontology provides a fine-grained, machine-readable description of what changes in OSM data over time.
This observable level is fully implemented and constitutes the current operational output of the pipeline. However, isolated transitions are inherently ambiguous. For instance, a tag change may reflect a correction, a refinement, or a real-world modification. To address this limitation, we specify an interpretative level – currently under implementation – that operates on sets of transitions rather than individual ones, aggregated along two axes. The first axe is Co-occurrence: transitions of the same type are grouped together when they appear massively across the study area within a given period. The second is Coherence: transitions of different types may be grouped when they share other characteristics, namely temporal coherence (simultaneous or close in time), spatial coherence (geographic clustering), contributor coherence (same OSM author or same changeset), or structural coherence (connected objects).

Interpretations are formalized through an ontology structured around four high-level categories: Representation refinement (semantic, attribute, or geometric updates that improve how features are modeled without correcting an error); Error Correction (fixes to incorrect category, attribute, or geometric information); real-world evolution (genuine infrastructure creations, deletions, or modifications); and ambiguous cases (transitions that cannot be interpreted with certainty). Inference rules map sets of observable transitions to interpretative categories, acknowledging that a single transition may admit multiple interpretations. Rules are organized by target interpretative category. For instance, for Representation refinement: IF the same tag modification occurs on a large number of features within a short time window AND across a wide spatial extent → THEN semantic refinement. And, for Correction: IF a feature is deleted AND a semantically equivalent feature is created nearby within the same changeset or time window → THEN correction by replacement.

The outputs of this pipeline are published as a Knowledge Graph (KG) following the Change Bridges approach introduced by the TSN-Change ontology [5]. Rather than representing OSM features themselves as a KG, as done in projects such as WorldKG [6], our KG exclusively describes the detected changes, linking each change node to the stable URIs of the corresponding OSM features. This design choice keeps the KG lightweight and focused on temporal dynamics, while remaining interoperable with existing geographic KGs.

Alongside the pipeline, we developed a dedicated web application to explore OSM editing trends over time and visually verify results by inspecting the lifecycle of selected changing features. The application operates directly on data derived from the full history planet file. At present, the application and the KG are two independent components. The former focuses on data exploration, the latter on change description and interpretation. However, the two are designed to converge: in future work, change annotations produced by the pipeline will be surfaced directly within the application's timeline, enabling users to interpret editing trends in light of their semantic classification and to better distinguish real-world evolution from data refinements.
The interface of the exploratory application is composed of various components: a geographic map displaying the cycling network at different points in time; interactive timelines showing the evolution of tag usage and frequencies over the full OSM history; an extended data analysis panel, including individual tag counts for each point in time, evolution of individual tags across time, tag transition counts and a Sankey diagram encoding the flow of semantic transitions between the defined categories (e.g., from “shared use path” to “dedicated cycleway”) across time. The map and the timeline are synchronized: selecting a time range on the timeline simultaneously filters the map. Figure 1 illustrates the general layout of the application and the synchronized map-timeline interface, while Figure 2 shows a Sankey diagram capturing the dominant transition flows within the cycling network of Paris.

Applied to the cycling network across the three study areas, the pipeline currently produces a KG encoding the detected and classified transitions at the observable level, covering the full OSM history of Lisbon, Grenoble, and Paris. The interpretative level is fully specified and under active implementation. Together, these contributions lay the groundwork for a robust and extensible framework for OSM change interpretation, with direct applications in urban monitoring and data quality assessment. Finally, as we are committed to an open science approach, we will publish the application’s code as open source and immerse the knowledge graphs into the Web of Data so that they are accessible to anyone wishing to explore them.