Pipeline Detection Using Uncertainty-Driven Machine Learning

Kurzfassung

The pipeline detection model developed in (Dasenbrock et al., 2021) has proven to be capable of generalizing well from training data originating from Great Britain to Northern Germany. The thesis at hand used a similar model but applied it to a more differing and heterogeneous region compared to the training data in order to test its generalizability: Spain. Insufficient IoU scores showed that the model is not able to satisfyingly detect pipeline pathways in Spain. It will be of great importance when applying the model to new regions that it is also specifically trained for the new regions. While the model is permanently applied to new regions and consequently more training data is added, the need for new training data will diminish with time. This is because the knowledge of the model will become broader, and the differences between new regions and the regions already shown to the model will likely decrease. To speed up this process and to train more sample efficient, the potential of an active learning approach was investigated.