Urban Heat Mitigation AI/ML

Geospatial, physics-informed AI/ML system to (1) map urban heat hotspots, (2) quantify drivers of urban heating, (3) model land surface temperature (LST) dynamics, and (4) generate + optimize cooling interventions.

acquire datasets ──> build drivers ──> train physics-informed model ──> simulate cooling scenarios ──> optimize + map
   (GEE/OSM)           (NDVI/NDBI...)       (monotonic XGBoost)          (cool roofs, greening...)

The production stack is live satellite data plus physics-constrained XGBoost:

cd C:\Users\yashasvi\Documents\Python\urban-heat-ai
python -m pip install -r requirements.txt

Real-data (one-time, do on Windows — browser opens for login):

# Set gee.project in config.yaml first, then authenticate + verify:
python scripts\gee_auth.py

Interactive UI (recommended) — pick any city/region and dates, then run the live analysis:

streamlit run dashboard\app.py

CLI — any place, any dates:

python scripts\run_pipeline.py                                  # AOI in config.yaml
python scripts\run_pipeline.py --city "Mumbai"                  # any city/state by name
python scripts\run_pipeline.py --city "Bengaluru" --start 2024-03-01 --end 2024-06-01
python scripts\run_pipeline.py --bbox 72.8,18.9,73.0,19.2       # manual box

Verify the data + predictions actually used:

python scripts\inspect_data.py     # cached raster stats + driver table + predictions

Tests:

python -m pytest -q

Outputs land in outputs/: lst_hotspots.png, driver_importance.png, validation_scatter.png, scenario_cooling.png, lst_map.html (open in browser), intervention_plan.csv, summary.json, saved model. The dashboard additionally exposes an OSM parcel plan as downloadable GeoJSON.

XGBoost uses hard monotonic constraints so increased vegetation, water, or albedo cannot produce physically reversed intervention estimates. It is deterministic, fast, and the only supported model path.

Reported on real Indian-city LST (New Delhi, Bengaluru):

The honest R² is what matters; the in-scene↔honest gap is the inherent ceiling of single-scene LST, verified un-closeable by tuning. What lifted it (all measured on cached tiles):

• Dropped ERA5 meteorology from the model — at city scale it's a near-constant position proxy that overfits (honest +0.06 to +0.15).
• Multi-scale neighbourhood context (_N 210 m + _NC 630 m) — matches LST's true spatial footprint (essential: removing it drops honest to 0.19).
• 2×2 quadrant CV for an unbiased honest estimate (vs a single unlucky corner).
• Landsat 8 + 9 merge and multi-year seasonal LST composite (time.lst_years) — denser, cleaner target.

Physics-informed drivers and monotonic constraints keep interventions physical.

Complete + tested end-to-end. The pipeline exports and reproject-matches all datasets to the Landsat grid (src/data/align.py), auto-scales exports to the GEE size limit, and is resilient to a failed side source. It needs Earth Engine authentication (python scripts/gee_auth.py), a GEE project id, and an AOI. One-page PDF/PNG reports are written to outputs/.