Minute-scale precipitation nowcasting (Phase 175.5).
| Module | Phase | Contents |
|---|---|---|
axial |
175.5.1.1 | Axial self-attention (row-then-column) — the MetNet-3 global-receptive-field primitive. |
metnet |
175.5.1.1 | MetNet-3-class nowcaster over stacked radar+satellite+topography+NWP context, lead-time conditioned, categorical rain-rate output (0–12 h, 1 km / 5 min). |
dgmr |
175.5.1.2 | DGMR-class conditional GAN: latent-conditioned generator (radar ensemble), spatial + temporal discriminators, hinge losses, rainfall-weighted grid-cell regulariser. |
convective |
175.5.1.3 | Convective-mode classifier — Steiner (1995) convective/stratiform separation + a trainable CNN emitting mixture weights (stratiform/convective/tropical/frontal). |
verification |
175.5.1.4 | CSI/POD/FAR + neighbourhood FSS at 0.1/1/5/10 mm h⁻¹ thresholds. |
Every model is real and trainable (overfit tests verify gradient flow); the FSS harness is validated against hand-computed contingency tables and demonstrates the neighbourhood metric crediting a one-cell displacement that pointwise CSI scores as a total miss.
bash
cd libs/gaia/nowcast && python3 -m pytest tests/ -q