The proposed solution involves comparing images of the same area taken at two different times. The system "learns" to recognise objects visible on maps, such as buildings, roads, forests or farmland, and automatically detects changes (including the addition or removal of objects in the space).
A combination of spatial analysis algorithms, computer vision and a modern artificial intelligence model was used for this purpose. It can analyse various types of data simultaneously, including colour aerial photos and LiDAR data describing the height of terrain and objects.
Studies have shown that combining these information sources significantly improves object recognition accuracy, especially in urban areas. The system is highly accurate in identifying buildings and roads, and performs even better in forested and agricultural areas. Importantly, the addition of LiDAR data significantly increased the efficiency of the entire process.
The detected changes are then classified as new objects, deleted objects or modified objects, and the results are presented as maps ready for further analysis in popular geographic information systems (GIS).
Although in the most complex situations, e.g. with strong shadows or obscured objects, user intervention is still necessary, the authors underline that the proposed solution is an important step towards faster, cheaper and more up-to-date map creation.
Dr Maciej Adamiak from the Institute of Urban Geography, Tourism Studies and Geoinformation at the Faculty of Geographical Sciences of the University of Lodz and the first author of the publication says:
– Our solution will bring the greatest benefits wherever quick information about the occurrence of changes in a given space is required, from crisis management, through climate action and the military, to construction.
He adds: – The method we have developed makes the map update process more accurate and delivers data much faster. This solution won't replace human operators, but it will significantly simplify their work. For the solution to be widely used by cartographic institutions and public administration, it needs to be integrated more efficiently with data sources and trained with a better semantic segmentation model based on one of the currently available base models. Then, production can begin.
Source: Adamiak Maciej, Będkowski Krzysztof, Nalej Marta, Ożadowicz Patryk, Pietruk Jacek, Wójcik Szymon: Bitemporal change detection for topographic map updates using panoptic segmentation of VNIR orthophotos and LiDAR data, GeoInformatica, vol. 30, 2026, Article number: 2, p. 1-41, DOI: 10.1007/s10707-025-00561-z
