SAAS River Forecasting Models

📁 Files

Model & Graph Structure

• best_model.pt: Pre-trained GNN model weights (PyTorch state dict)
• hja_edge_index.npz: Graph edge connectivity matrix for H.J. Andrews sites
• hja_graph.gpickle: Complete graph structure with node features (NetworkX format)

Training & Evaluation Data

• static_vars_pivot.csv: Static watershed characteristics for all sites in pivot format
• degrees.parquet: Node degree statistics for graph analysis
• obs_pivot.csv: Observational data used for training and evaluation in pivot format

RGCN Predictions

• train_val_predictions_day1.csv: RGCN predictions for 1-day ahead forecasting horizon
• train_val_predictions_day2.csv: RGCN predictions for 2-day ahead forecasting horizon
• train_val_predictions_day3.csv: RGCN predictions for 3-day ahead forecasting horizon
• window_split_map.csv: Mapping of window indices to train/val split assignments used in RGCN evaluation

Reference Data

• nhd_id_stream_order_permanence.csv: Stream order and perennial/intermittent classifications from USGS NHD

🚀 Quick Usage

💻 Computational Requirements

• GPU Required: NVIDIA RTX A6000 or equivalent (48GB VRAM) for training
• Training Time: ~4 hours
• Inference: CPU compatible (slower)
• Dependencies: PyTorch, PyTorch Geometric, NetworkX, pandas, numpy

🏢 Affiliations

• Student Association for Applied Statistics (SAAS), University of California, Berkeley
• United States Geological Survey (USGS)