Artificial intelligence in surveying and geoinformatics

Artificial intelligence (AI) is now ubiquitous - including in surveying and geoinformatics applications. AI methods help with the analysis of large and complex data sets, such as satellite and aerial images.

Teaching the basics of AI is a central component of all surveying degree programmes at HFT Stuttgart. The concepts and methods of supervised and unsupervised learning form the basis for this. In unsupervised learning, methods group data according to similarities, for example similar spectral properties. This can be used to find areas in image data that are covered with similar materials. In supervised learning, models are trained to find already known classes in the data. To enable this, the method is shown examples on the basis of which it then learns its task. Such methods are used in the classification of image and other data. Traditional methods such as Maximum Likelihood, k-Nearest Neighbour and Random Forest are taught and used, as well as modern neural networks. The aim of each course is to teach both the theoretical foundations and practical applications.

For the application of AI methods, especially for the training of neural networks, powerful computers with suitable graphics cards are required. The appropriate infrastructure is made available to students in the GeoAI lab at HFT Stuttgart.
 

As an example of the successful teaching of AI methods in the field of surveying, we would like to cite the following research contribution, which was developed in collaboration with Stadtmessungsamt Stuttgart (City Surveying Office).

As part of the Remote Sensing Studio module in the Photogrammetry and Geoinformatics Master's programme, different methods for mapping vegetation in urban areas were compared. The study used high-resolution aerial image data from Stadtmessungsamt  Stuttgart, which kindly provided us with the data for this project. As part of the module, the image data was independently annotated by the students (to generate sample data) and the joint data set was made available to all three projects. It turned out that the semantic segmentation and object recognition approaches work very well for recognising low and high vegetation. Further analyses also showed that the presence of a near-infrared band is not very important for the detection of vegetation and that the trained models also achieve good results on other aerial images. The results of this project will be presented at the annual conference of the German Society for Photogrammetry, Remote Sensing and Geoinformation in Mainz in 2026.