MT_GCNN

The source code of the manuscript entitle "Move and remove: Multi-task learning for building simplification in vector maps with a graph convolutional neural network"

Requirements

• python 3.7
• pytorch 1.11.0
• torch-geometric 2.0.3

Data description

The building graph features and labels are stored in the directory data/input/ with numpy .npy file. The description for each columan of the file is as follows:

• Column 0: osm id of the building corresponding to the vertice
• Column 1: vertex id of the vertice
• Column 2: longitude of the vertice
• Column 3: latitude of the vertice
• Column 4: normalized longitude of the vertice
• Column 5: normalized latitude of the vertice
• Column 6: turning angle of the vertice
• Column 7: convexity of the vertice
• Column 8: preceeding edge length of the vertice
• Column 9: succeeding edge length of the vertice
• Column 10: the removal label
• Column 11: the movement label along the preceeding edge
• Column 12: the movement label along the succeeding edge

Usage

Config the input directory and hyperparameters in main.py and run it. In case of understanding the proposed MT_GCNN model, check it in models.py with the class BuildingGenModel.

• data/input: the directory where vertex feature and edge adjacency files are input to the model
• data/output: the directory where the trained model (Bldgs_Gen_64_1.pkl) and the predicted vertex labels (the point file: Bldgs_Gen_prediction.shp) are stored
• Note: to get final simplified buildings, please use the reconstruct_polygons function in utils.py to reconstruct polygons based on the output point file Bldgs_Gen_prediction.shp

Citation

@article{zhou2023move,
  title={Move and remove: Multi-task learning for building simplification in vector maps with a graph convolutional neural network},
  author={Zhou, Zhiyong and Fu, Cheng and Weibel, Robert},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={202},
  pages={205--218},
  year={2023},
  publisher={Elsevier}
}