• Post category:Articles
  • Post published:14.12.2023

Artificial intelligence improves the National Land Survey’s topographic data accuracy

In this article, the researchers will highlight the findings of the AI4TDB project which aimed to advance the accuracy, automate object collection, and ensure the up-to-dateness of spatial data within the National Land Survey’s Topographic Database. The results are attracting international interest as the use of artificial intelligence to improve the accuracy of topographic data production is a topic of discussion.

You are currently viewing Artificial intelligence improves the National Land Survey’s topographic data accuracy
Overview of Oulu and Jyväskylä test areas: true orthophotos with AI-detected buildings (red).

The National Land Survey of Finland (NLS) initiated the open release of its topographic database (TDB) in 2012, amplifying its use in various applications. The demand for precise spatial information surged due to innovations like location-based services, mobile mapping, and autonomous driving. To meet this escalating need, refining the accuracy of the NLS topographic data became paramount.

The TDB comprises different vectors such as buildings, roads, and watercourses. However, owing to historical factors, certain vectors may exhibit inaccuracies. The evolution to digital photogrammetry workstations and subsequent advancements in digital cameras drastically improved aerial image resolution, accompanied by robust software and hardware facilitating precise calculations. Internal projects like ATMU and KMTK 3D identified location inaccuracies in TDB building vectors, emphasizing the need for improvement, leading to the inception of the AI4TDB project.

AI4TDB, standing for ‘Artificial Intelligence for Topographic Database Accuracy Enhancement,’ co-funded by the Ministry of Agriculture and Forestry of Finland and the NLS, aimed to leverage AI to refine the precision of the database, focusing on buildings and watercourses. The project used deep learning solutions developed by the ATMU project (buildings obtained from true orthophotos) to improve the TDB vector location accuracy levels. Additionally, the project underwent further refinement for the AI model by training it with new datasets, significantly enhancing its performance.

For small watercourses, traditional mapping methods faced challenges, especially in densely forested areas. Prior watercourse detection projects using deep learning methods and Lidar-DEM revealed shortcomings in detecting smaller water bodies that connect small watercourses. The AI4TDB project aimed to rectify this by integrating ponds into watercourse networks to ensure comprehensive and accurate data.

The project’s objectives included refining detection methods, rectifying location disparities in buildings, and developing AI solutions for identifying elements within watercourse networks. Ultimately, these efforts aimed to produce higher-quality map products while ensuring the database’s accuracy and up-to-dateness.

The watercourse recognition  a) topographic position index background layer, b) watercourse labels, c) model predictions


In this project, for the topic ‘buildings’, the datasets: true orthophotos, DEMs, and DSMs have been used in the UNet model, previously trained, for obtaining ‘AI-detected building vectors’. The AI model generated reference vectors, aiding in identifying and correcting location errors in the TDB buildings. Throughout the project, data from 11 regions underwent testing, covering a total area of 1628 km2. Computing was done on a high-performance CSC platform.

For the topic ’watercourses’, ditches, natural streams, and their connecting ponds were recognized using UNet convolutional neural network (CNN) with teaching areas of three different terrain types. Digital elevation models from LIDAR point clouds, near-infrared orthophotos, and regular orthophotos were used as input data. Important emphasis was given to careful in-house digitization of watercourse features. This appeared crucial as offsets of only a few pixels from the channel centreline were found to cause a significant decrease in recognition accuracy. The digitization was made from orthophotos, hill shading, and topographic position index.


During the project, we have handled many challenges. With regard to buildings, the AI4TDB project identified challenges in the quality of data sources (such as true orthophotos and Lidar data), the environmental context (urban or forested areas), and the AI model’s adaptability to new datasets.  Manual intervention, validated by a two-phase quality control process, was crucial despite intelligent strategies. Anticipating reduced manual efforts, comprehensive training of the AI model with varied data sources is expected. Experiments with different types of input data revealed the limitations of using true orthophotos in forest areas and highlighted the advantages of Lidar data, proposing a hybrid approach for optimal results in future endeavors.

Recognizing watercourse types posed challenges for CNN recognition due to varying terrains. Clear ditches were easiest to extract (86-95% detected), and even filled ones were detected. Natural streams and ponds proved harder due to their unclear edges and diverse physical forms. Despite this, combined, these watercourse features form a comprehensive network. Further enhancements can be achieved with vector operations improving connectivity and integrity.


Close collaboration between mapping and AI teams was crucial. The AI team’s initial results were examined by mapping teams, fostering mutual understanding and enhancing knowledge exchange. As members of the National Land Survey (NLS), we recognize the challenges within the TDB and are committed to gradual enhancement. The insights gained hold significant value for stakeholders, companies, and researchers, promising notable improvements in TDB data quality and substantial benefits across sectors in the near future.

The NLS mapping teams have identified new applications for AI-generated buildings: locating missed buildings and removing demolished ones from the TDB vectors. Initial tests suggested hundreds of missed and numerous demolished buildings in a 144 km2 urban area. Expert evaluation of relocated and rotated TDB buildings is ongoing in the project’s final phase, guiding future applications. Simultaneously, ongoing development aims to create comprehensive watercourse networks for future integration into the NLS TDB.


Lingli Zhu, D.Sc. (Tech.) AI group leader, Chief Specialist, National Land Survey of Finland
Jere Raninen, IT Expert, National Land Survey of Finland
Emilia Hattula, IT Expert, National Land Survey of Finland
Pyry Kettunen, D.Sc. (Tech.), Research Manager, Geospatial Research Institute, National Land Survey of Finland
Christian Koski, Research Scientist, Geospatial Research Institute, National Land Survey of Finland
Anssi Jussila, Research Scientist, Geospatial Research Institute, National Land Survey of Finland
Juha Oksanen, Prof., Head Of Department, Geospatial Research Institute, National Land Survey of Finland

At the Location Innovation Hub, we have collected several location-based AI methods developed by top researchers and our partner companies. We provide AI models, supporting training data, research publications and other documentation. If you are interested in AI solutions around geodata, do not hesitate to contact us.