The 3rd Forum GEO.KI will take place from 1–2 October 2026 in Frankfurt am Main. The event brings together people from public authorities, research, universities, and industry to exchange ideas and share practical examples of how artificial intelligence can be used with geospatial data. This year’s program covers topics such as GeoAI methods, Earth observation, infrastructure, and real-world applications. HeiGIT will present a poster on a GeoAI approach for validating crowdsourced land-use edits in OpenStreetMap using satellite imagery and deep learning-based change detection.
Poster: Temporal validation of crowdsourced land-use edits in OpenStreetMap using deep learning-based satellite change detection
Presenters: Clemens Langer, Mohammed Rizwankhan
OpenStreetMap (OSM) is constantly being updated by a global community, making it an incredibly valuable source of land use information, especially for a country like Germany where OSM landuse tags comparable to CORINE classes are more than 80% complete and are frequently updated by the community. However, because these updates are crowd-sourced, it can be difficult to assess how accurate or consistent they are over time—especially at scale. In this work, we present a practical GeoAI approach to evaluate OSM land use changes using satellite imagery.
Our method links multi-temporal satellite data with recent OSM changes, such as feature creations, updates, and deletions. These OSM edits are fetched from ohsomeDB, which maintains a comprehensive records of major and minor OSM element versions, including geometry, enriches with attributes like area and centroid, valid time ranges, and country codes. We spatially organize these edits using the H3 grid system and, for each affected area, retrieve two PlanetScope images (3 m resolution) taken roughly one year apart. These image pairs are then analyzed using a change detection model (ChangeStar from Pytorch), which we retained to work effectively with medium-resolution imagery we have from PlanetScope. The model produces a probability map showing where changes are likely to have occurred.
We then compare these changes in probability map with the corresponding OSM features and calculate an “adequacy score” that reflects how well the mapped edits align with actual changes on the ground. This makes it possible to flag individual edits that are well-supported, uncertain, or potentially incorrect.
A key part of this work is improving the performance of change detection models on 3 m PlanetScope imagery, which is more widely available but also more challenging than very high-resolution data. Overall, our approach shows how combining volunteered geographic data with Earth observation and Deep learning models can support large-scale quality assessment. It has practical applications for improving map reliability, guiding validation efforts, and supporting use cases in urban development, environmental monitoring, and land management.