There are several 3D city models available openly, worldwide. These models are used in various applications, from which many expects a homogeneous Level of Detail (LoD). Validating the accuracy of the LoD of a model requires the inference its LoD class and its conformance to the real-world object. This process quickly becomes infeasible for large models when done manually. Yet there is no automatic method for LoD inference and validation. Therefore the thesis proposes a method to automatically infer the geometric LoD (LoD0-2.3) in 3D city models.

A central aspect of this work is the use of machine-learning to classify building models based on their LoD. It follows the assumption that a process is possible where a classifier trained in a synthetic 3D city model containing all LoD classes, and applied in real city model. Therefore ten geometry measures (features) are computed from the objects and tested with six classification algorithms. The six experiments the transferability of a classifier from the synthetic city model to the real one, multi-class (LoD0-2.3) and binary (LoD2 or not) classification, and the effect of LoD class imbalance by introducing various amounts of LoD1 objects into the LoD2 model. Furthermore, by using a point cloud as ground truth, this explored the possibility of validating the inferred LoD classes.

The results indicate that the classifier is not transferable to the real data set when trained on the synthetic city model, which is probably due to the significant difference in object shapes between the two models. Binary classification outperforms the multi-class case and it is favourable for LoD validation where the main question is whether the model conforms the stated LoD or not. Finally, class-imbalance can reduce the classification with as much as 20%.