GALATEA

Graph-based Analysis and Learning with Aquatic Telemetry for Environmental Assessment

Graph neural networks for predicting biochemical oxygen demand (BOD) in freshwater from low-cost sensor data. GALATEA pairs graph-based matrix factorization, which imputes the readings low-cost sensors miss, with a GNN over the completed spatiotemporal graph. It predicts BOD at 0.27 mg/L MAE, a 22.85% improvement over existing low-cost models, running on a $385 sensor suite in place of $20K+ commercial monitoring hardware.

1st place, Maryland, 2025 Stockholm Junior Water Prize; represented Maryland at the US national competition.

How it works

1. Graph creation (model/graph_creation/): water-quality readings become a spatiotemporal graph, with sensor measurements as node features and spatial/temporal relations as edge attributes (BOD_graph_creation.ipynb).
2. Imputation (model/GMF_GNN/): GNN-assisted matrix factorization fills missing environmental variables, in two message-passing variants: GCN and edge-aware GATv2 attention (GMFGNN_GCN.py, GMFGNN_GATv2.py).
3. Prediction (model/GNN/): a GNN over the graph predicts BOD, trained with neighbor sampling, GraphNorm, and mixed precision (GNN_GCN.py).
4. Model selection: sweep scripts over layer counts, activations, edge dimensions, and neighborhood sizes (model/*/**.sh), compared in model/compare_stats.ipynb.

Stack

PyTorch, PyTorch Geometric (GATv2Conv, GCNConv, GraphNorm, NeighborLoader), torcheval, scipy.

Poster