Residential Building Type Prediction Using OpenStreetMap Data in Germany

Kurzfassung

In crisis scenarios, distinguishing between building types - particularly residential and non-residential buildings - is crucial for the reliability of the digital twin. The underlying data is sourced from OpenStreetMap; however, for over 67 % of buildings in Germany lack a specific building type. Existing methods for predicting residential buildings were developed for the United States and are not transferable to Germany due to geographical differences. To develop a valid and practical foundation for the digital twin, a method for predicting building types in Germany must be developed. Therefore Prediction Models (Logistic Regression, Decision Tree, Random Forest) and Logic Prediction are used which use several binary predictor variables as well as a numeric predictor variable. The results are evaluated with precision and recall, the regions are chosen for a cross-scale prediction and a zero-shot transfer. The results show that machine learning algorithms are suitable for complex, non linear problems but the accuracy depends on the completeness of the training dataset.