Measuring urban environmental data with AI
Data on urban geometry (e.g. building facades) and environmental data (e.g. air temperature) form the basis for recording the local urban climate in detail and deriving measures for improved climate adaptation.
Recording and interpreting urban geometry with AI
With the aim of creating a 3D view of the city of Freiburg in which object classes (houses, trees, etc.) are recognised and labelled using deep neural network techniques, the following steps were taken in this work package: An artificial neural network (ANN) that can work directly on 3D data was trained on the basis of existing so-called mobile mapping data generated in the project, such as georeferenced image and laser scan data from inspections or aerial surveys of the city of Freiburg. The neural network can recognise object classes in so-called 3D point clouds and assign a label (also known as semantic segmentation). In future, this will make it much easier to classify new inspection data for the classes trained in the project. For the required training and validation data, point clouds of sub-areas of the city of Freiburg were first annotated by humans (i.e. recognised and given a corresponding label) (see Figure 1). In addition, AI-based preloaders for 3D visualisation were researched, which enable the collected and processed data to be visualised in a browser application (similar to the existing 3D model of the city of Freiburg).
Improving local weather recording and environmental monitoring with the help of AI
The analysis of spatial and temporal weather patterns at a local level in cities was also improved by harmonizing and validating various data sources using artificial intelligence methods. The geographical and climatic variability of Freiburg's urban area and the surrounding area can be recorded using existing measuring stations and with the help of a specially constructed urban climate measurement network, which was set up in cooperation with the ERC grant “urbisphere”. The measurement data from currently 42 stations serve as the basis for a variety of research projects in the fields of meteorology, hydrology and urban planning. The environmental data collected, such as air temperature, precipitation, wind speed and global radiation, undergoes automated quality control and will be made available to the public in real time in the future. In addition, an AI-based gap-filling method was developed that can be applied to meteorological measurement networks worldwide on a local and regional scale. Based on machine learning methods, this procedure can be used to accurately fill existing gaps in the time series of individual stations with the aid of the remaining stations and, using the time series obtained in this way for each station, to calculate climatological, statistical values, such as the mean air temperature at each location, as well as the number of climatological knowledge days, such as the number of tropical nights per year for all stations.