The Advanced Technology for National Topographic Map Updating (ATMU) project was one of the first artificial intelligence (AI) projects in topographic data production of the National Land Survey of Finland (NLS). The goal of the project was to reduce the amount of manual and routine work in topographic data production, increase the accuracy and up-to-dateness of the data, improve efficiency and save costs of the NLS, enhance the map updating technology automation level in NLS, and promote the use of AI technology for national mapping products.
The ATMU project developed deep learning solutions for feature extraction and updating, focusing on buildings, roads, and watercourses. The project focused on two aspects: method development for the use of AI, and sharing knowledge and experience with society (Fig. 1). In the method development, special attention was paid to i) high-quality and large-quantity training data from diverse environments, ii) different neural networks and one with the best performance was selected to be used in the project. Social impact is an important measurement of a project’s success. During the project, public seminars were organized every half a year to share our knowledge and experience in collaboration with Geoforum Finland, Cartographic Society of Finland, and the ICA’s Working Group on the Digital Transformation of National Mapping Agencies. Besides, at the end of the project, high-quality training data was released to the public. In addition, as byproducts of the ATMU project, more than 100,000km2 true orthophotos were produced during the project. Through the two-year effort of the ATMU team, the project achieved great success and was fruitful.
Figure 1. An overview of the tasks in the ATMU project
The ATMU project had development tasks on object detection and change recognition. Object detection is an important step in information acquisition automation from imagery. In the ATMU project, object detection focused on three types of objects: buildings, roads, and watercourses. For building detection, true orthophotos, building vectors, DEMs, and DSMs were utilized to train the UNet model. The outputs were the building outlines from the UNet. In road detection, either orthophotos or true orthophotos with road vectors were used to create training data. RoadVecNet (released in 2021) was employed to detect the roads. Along with multitask learning techniques, both road surfaces and edges can be obtained from the neural network. In watercourse detection, high-quality watercourse vectors were manually digitized from DEM-derived terrain visualizations and orthophotos. Clarity classes were also defined for digitized watercourses in order to enhance detection and vectors with 1.5m buffers were used to create the training data. Together with self-created 0.5m Lidar-DEM, the UNet model was trained. The segmented watercourses were output from the neural network. Table 1 summarises the deep learning solutions for different object detection in the project and Fig. 2 shows the results from neural networks.
Figure 2. Results from neural networks for buildings (UNet), roads (RoadVecNet utilizing multitask learning technique), and hydrographic features (UNet).
Besides object detection, change recognition for buildings and roads was also studied. Different neural networks such as DSAMNet, MDESNet, Changer, NestNet, and NestNet2 were tested. As a result, NestNet2 performed the best in predicting the changes in buildings and roads from true orthophotos captured in different years.
NLS is a national government agency in Finland providing information and services about the earth, sharing data and map products with the public, and researching and developing advanced technologies and methods. In 2012, the NLS was one of the European frontrunners by releasing the topographic database as open data. Under the open spirit in the NLS, the ATMU project shared the AI knowledge and experience from the project by organizing open seminars once every half a year. During the past two years, the project has hosted four open seminars. More than 300 international participants have participated in the events. Among these, one event had 150 registered people. In each open seminar, the experiment on AI methods, experiences, and lessons was shared and openly discussed. Besides the open seminars, we released high-quality training data to the public. Thus, our work was beneficial for the geospatial community to boost AI development. The social impact is an important measurement of the success of a project. Besides open seminars and open training data releasing, the ATMU project was also introduced in many workshops, conferences, GIM- International magazine, and EuroGeographics’ 2022 annual report. Three Master’s theses were also completed in the project.
In December 2020, the Advanced Technology for National topographic Map Updating (ATMU) project was selected as one of the funded projects (so-called “Robo funding”) by the Ministry of Finance. It was a two-year project for the period of 1.1.2021-31.12.2022, with a funding of 399,200 €. The ATMU project has developed methods for detecting buildings and roads from aerial imagery (including true orthophotos and orthophotos) and detecting watercourses from Lidar-DEM. By the end of the project, more than 100,000km2 true orthophotos were produced, and the UNet model for building detection had been trained with datasets selected from data from different areas of Finland covering more than 60,000 km2. According to the NLS expert evaluation, it has reached an accuracy level of up to 97.9% when compared to different reference data. It is now available for practical use in map production. The option of integrating the ATMU building model with the new topographic database system is currently under discussion. The AI model (RoadVecNet) for road detection likewise achieved a great result and further application is under consideration. Meanwhile, the preliminary results for the complex watercourse detection using UNet provide a good basis for further research. In addition, expert evaluation of building and road change recognition had been implemented. The results showed that approximately 96% of changes from roads and buildings have been found from true orthophotos using the NestNet2 model. The change detection results can be used as pointers for the operators so that they do not need to check the changed areas across the entire image. This will significantly reduce the amount of manual work. Further application of the change detection method might need to be discussed in the coming years. The ATMU project also produced high social impact by organizing open seminars to share knowledge and experiences, releasing high-quality training data to the public to boost AI development in the geospatial field, and also introducing the project in many workshops and conferences nationally and internationally.
The AI method development and application in the production of the NLS continue after the ATMU project. The new project AI4TDB utilizes the AI models developed in the ATMU project to improve the accuracy of vectors in the topographic database for buildings and watercourses.
The high-quality training data for building detection from the ATMU project has been released to the public on 1st November 2022. The training data included true orthophotos and their corresponding labels. It covered an area of about 50km2. The scenes include urban, suburban, and rural areas. You can download them from the this link.