Summary:
Distribution-agnostic graph transformer architecture.
The project aimed at predicting landslide extents on a highly right-skewed dataset.
It uses a non-standard loss function, the Squared Error Relevance Area (SERA), and the Mean Squared Error (MSE) as performance metric.
Research article available here
Reproducibility:
Within the environment created from pythonenv/environment.yaml, run run.py. This will execute the model tenfold, once for each of the ten deciles used as test set.
Alternatively, it is possible to run the model for one single test set by specifying the desired test set decile in configs/SU_params.json ("dataset":"pickle") and running main.py.
(Optional) Run pred_visualization.ipynb to visualize the predictions.
Before running, please extract the gpkg contained indata/custom/Wenchuan_data_final.zip into data/custom/.
Note:
data_prep.ipynb is provided as reference to share the preparation of the original gpkg dataset as the model input.