Geographic transferability of building function models

Kurzfassung

We study the geographic transferability of U-Nets trained for building type classification. The models assign functional classes, such as residential or non-residential to individual buildings. This extends the common building extraction task by providing more in depth insight. Since training and testing data is generally scarce, it is of fundamental importance to study transferability of models between different areas. To address this, we collected a dataset covering six countries across the globe. We paired Planet Basemap images with OpenStreetMap (OSM) labels and manually cleaned the dataset. We trained semantic segmentation models on various data splits to predict residential and non-residential buildings. The presented results suggest there is no clear benefit from training or fine-tuning models locally, compared to combining the available data and training a global model. Furthermore, we find stacking footprints on top of the input images improves both, the models’ F1 scores and across-region transferability.