Deep Learning in the Built Environment for Large-Scale Mapping and Detailed Land Use

Highly detailed, large-scale land-use monitoring so far in Belgium has been absent. Lots of data is available such as satellite and aerial imagery but also point cloud data from LiDAR. However, the handling of that data is either a manual interpretation task or a heavily supervised machine-learning method is used. The manual labeling of training data for machine-learning networks is a common approach, but highly intensive and costly. Therefore, this PhD research will look into new weakly and semi-supervised machine-learning methods requiring only a limited set of labeled data. Exploiting other data sources as training data, such as the GRB or land-use maps, will contribute to the ease of use of machine-learning networks. This combination of technologies will allow for much quicker processing and interpretation of various forms of remotely sensed data. It will prove its use in highly detailed (cm-level) observation of land use and detection of building change, but it can also be used on a terrestrial level for progress monitoring of construction sites.