Next-day forest-fire risk for Uttarakhand, India — backtest-only research demo. Predicts, for a grid cell, whether a fire will be detected there tomorrow, using fire history, weather, and human-proximity/land-use features (roads, settlements, land cover).
The hypothesis this exists to test: in a fire regime that is 90–95% human-caused, do human-activity and land-use-adjacency features contribute more predictive lift than weather alone?
Planning complete, build not yet started. Full spec (requirements, architecture, screens, API/data contracts, execution plan) exists as local working docs, not published here.
The four files below are carried over from an earlier, unrelated US-focused
project, kept only for their patterns — nothing in the eventual src/
package will import them:
firepulse.py/test_firepulse.py— leakage-safe panel construction and its test style (label is always day d+1, no feature sees the future)scripts/download_era5.py— resumable, idempotent chunked download pattern for weather reanalysis via the CDS APIscripts/train_forecast.py— staged ablation runner pattern
These will be retired once their patterns are ported into this project's own structure.
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txtMIT — see LICENSE.