Overview
Global wildfire models struggle to represent local conditions: fire is a regional phenomenon. For my thesis I designed a Local Area Modeling (LAM) approach on top of the FireCastNet GraphCast-style architecture — graph neural networks over icosphere meshes that forecast burned area from climate-cube inputs.
The code was merged into the upstream project and lives in the lam and lam-visual branches of FireCastNet:
lam · lam-visual.
What I built
Regional modeling needs meshes that concentrate resolution where fire happens. I built a pipeline that constructs per-region icosphere meshes (GFED regions such as BONA, CEAM, CEAS) with PyMesh in Docker, and data pipelines that slice the SeasFire climate cubes into region-local inputs (mslp, tp, vpd, sst, t2m_mean, ssrd, ndvi, population density) with gwis_ba burned area as the classification target.
The model uses the GraphCast encoder–processor–decoder stack with mesh-node GRU processors. Training is config-driven through the Lightning CLI with cell-area-weighted losses (L1/MSE/Huber) so grid-cell area is respected in the objective, AUPRC-based checkpointing and Weights & Biases logging. Baselines include GRU, ConvGRU, ConvLSTM, UTAE and custom GAT/GraphConv/GraphMsgPassing models.
Technical highlights
- Regional icosphere mesh construction for local-scale graph modeling
- GraphCast-style spatiotemporal GNN: encoder–processor–decoder with GRU processor
- Geospatially-aware cell-area-weighted losses
- Full Lightning CLI experiment stack with W&B tracking and AUPRC metrics
Outcome
The thesis results were contributed upstream; the lam branch carries the modeling work and lam-visual the visualization of its outputs.