Rethink Geographical Generalizability with Unsupervised Self-Attention Model Ensemble: A Case Study of OpenStreetMap Missing Building Detection in Africa
In this work, we proposed Geographical Weighted Model Ensemble (GWME), an unsupervised model ensemble method to improve the geographical generalizability of GeoAI models, with a case study of cross-country OpenStreetMap (OSM) missing building detection in sub-Saharan Africa. Moreover, we compare four unsupervised model ensemble weighting strategies: 1) Average weighting, 2) Image similarity weighting, 3) Geographical distance weighting, and 4) Self-attention-based weighting. One can find the source code as follows.
docker build . -t gwme:<TAG>
docker run -it --gpus all --name gwme -p 8888:8888 -p 6006:6006 --mount type=bind,source="$(pwd)",target=/app gwme:<TAG>
jupyter notebook --ip=0.0.0.0 --no-browser --allow-root --debug
- Train: ./GWME/model_main_tf2.py
- Inference: ./GWME/inference_main.py, ./GWME/prediction_to_geojson.py
- Export: ./exporter_main_v2.py
- Evaluation: ./GWME/evaluation_geojson.py
- ./GWME/calculate_weights.py
- ./GWME/vit_representations.py
- ./GWME/merge_image_patch.py
- ./GWME/ensemble.py
- predictions
- Probing_ViTs: pre-trained ViT
- sample_images: satellite image tiles from reference area and target area
- ViT_sample_images: image candidates for calculating attention weights
- all_weights.json: tile-wise weights dictionary
Hao Li. Jiapan Wang, Johann Maximilian Zollner, Gengchen Mai, Ni Lao, and Martin Werner. 2023. Rethink Geographical Generalizability with Unsupervised Self-Attention Model Ensemble: A Case Study of OpenStreetMap Missing Building Detection in Africa. In Proceedings of the 31th International Conference on Advances in Geographic Information Systems (ACM SIGSPATIAL '23)
Hao Li: hao_bgd.li@tum.de
Hao Li is with the Technische Universität München, Professur für Big Geospatial Data Management, Lise-Meitner-Str. 9, 85521 Ottobrunn