Spatio-Temporal Graph Convolutional Networks for Shoreline Prediction

This repository contains code for GraphShore, a shoreline prediction model using spatio-temporal graph convolutional networks (GCNs). It incorporates the DSTD-GCN and the STSGCN model. The GraphShore model is introduced in this paper.

Folder Structure

• runner.ipynb: Main notebook for training, testing, and visualization.
• checkpoints/: Model checkpoints for each site.
• configs/: Configuration files for different models and sites.
• datasets/: Shoreline and wave datasets for each site.
• engine/: Prediction engine to train and predict.
• model/: Model implementations (dstdgcn, dstdgcn_nospatial,stsgcn) and layers.
• results/: Output predictions, targets, and visualizations.
• utils/: Utility functions for data processing, logging, visualization, and loss functions.

Getting Started

1. Install Dependencies

   • Python 3.10+
   • Required packages: numpy, torch, pandas, geopandas, matplotlib, optuna, tqdm, rasterio
   • Install with pip:

     pip install -r requirements.txt

2. Prepare Data

   • Place shoreline and wave datasets in the datasets/ folder following the same formatting as examples, organized by site.

3. Run Experiments

   • Modify and run runner.ipynb for model training and testing.
   • Modify and run ablation_analysis.ipynb to compare the performance of spatio-temporal and temporal-only models.
   • Modify and run model_interprete.ipynb to compare the learnt adjacency matrix with parameters used in physics-informed models.

Model framework

Acknowledgements

The model is built on the DSTD-GCN and the STSGCN model. We really thank the authors for sharing their wonderful works.