PyroAdapt lifts California wildfire prediction, boosting extreme-fire recall by 39.7 points
UIUC-CS · hf · 2026-09-29
A UIUC team introduces PyroAdapt, a pretrain–retrieve–rank framework for adapting wildfire occurrence models to spatial heterogeneity and temporal distribution shift. It retrieves historically similar locations and fine-tunes a pretrained model with risk-ranking objectives conditioned on terrain, ecoregion embeddings, and local fire rates.
Key results over California (666 grid cells)
- Daily average precision rises from 21.62% (focal fine-tuning) to 24.35–24.57%; Top5% recall from 18.70% to 22.11–22.79%
- Under a fixed daily budget of 34 cells (5% area), selective ranking captures 344 additional positive cell-days
- For fires in the top 5%/10%/20% of dry-matter consumption, recall gains are 39.70/28.18/20.50 percentage points
- Rolling evaluations at Yosemite show ranking gains persist under temporal shift
More from Research
- Why Sampled Softmax Speeds Training 1.7x — and How It Systematically Undertrains the Tail — tokenbender · 2026-09-29
- Why distillation beats vanilla supervised learning: it passes the full probability vector — khademinori · 2026-09-29
- AI Simulated 100 Papers on LZ Dark Matter Anomaly, Compared Against 82 Real arXiv Papers — skdh · 2026-09-29
- SUMI distillation study reports 35% SSIM gain on degraded PCCT data, but clinical benefit remains unproven — maier_ak · 2026-09-29
- Apollo Research: Models in Coding Evals Favor Graders Over Users, Reward-Seeking Grows With RL — burny_tech · 2026-09-29
- Late-layer neurons in Qwen act like on-off switches, unlike Olmo — Sauers_ · 2026-09-29