AI-powered Urban Heat Island prediction, monitoring and mitigation recommendation system.
Mumbai's built-up wards run measurably hotter than its green and coastal ones. This project quantifies that gap from open satellite data, explains why each neighbourhood is hot, simulates what would happen if you intervened (plant trees, coat roofs, add water bodies), and exposes the whole thing through a map dashboard and a natural-language copilot that urban planners can actually query.
Final-year major project. Built entirely on free tiers and open data — no paid services.
Live: urbanheat-mumbai.vercel.app — backend at urbanheat-api.onrender.com (free tier: the backend sleeps after 15 min idle, ~1 min to wake on first load).
| Capability | How |
|---|---|
| Measure heat | Land Surface Temperature from Landsat 8/9 thermal bands, dry-season composites |
| Explain heat | XGBoost/LightGBM trained on vegetation, built-up density, albedo, land cover, population; SHAP attributes each cell's temperature to its causes |
| Rank priorities | Heat Vulnerability Index = heat exposure × population × lack of green cover, aggregated per BMC ward |
| Simulate interventions | Digital twin: change a cell's features → re-predict → ΔLST map, with clamping disclosed when a scenario pushes past the model's training envelope |
| Recommend actions | Planning agent ranks interventions by ΔLST × population affected — no cost axis, since no cited cost-per-area figure exists yet (api-reference.md) |
| Monitor conditions | Scheduled job watches forecasts and raises heatwave alerts |
| Converse | RAG copilot answers planner questions over city data + policy documents |
 Heat map — every 200m cell, coloured by surface temperature |
 Click any cell for its real SHAP attribution |
 Scenario simulator — re-run the model with an intervention applied |
 Copilot — an agent calling the same tools, cited and grounded |
Full walkthrough script: docs/demo.md
Four LangGraph agents (Planning, Digital Twin, Monitoring, Copilot) sit on top of an ML prediction service, a GIS processing service and a scenario engine, all served by FastAPI to a React dashboard.
Full component and data-flow diagrams: docs/architecture.md
Data Google Earth Engine (Landsat, Sentinel-2, ESA WorldCover, WorldPop, SRTM) · Open-Meteo · OpenStreetMap/OSMnx ML scikit-learn · XGBoost (training) / xgboost-cpu (deployed) · LightGBM (training-time comparison) · SHAP Backend Python 3.12 · FastAPI · LangGraph · Gemini Flash · ChromaDB + Gemini embeddings Frontend React · TypeScript · Vite · MUI · react-leaflet · Recharts Infra GitHub Actions · Vercel · Render (Docker) · Supabase (Postgres + Auth, RLS)
Why each of these: docs/decisions/
Phase 6 complete — deployed and live (see Live link above). Data pipeline, ML model, backend API, four LangGraph agents, RAG copilot, React dashboard, auth, and public deployment are all built and running. Phase 7 (report, polish, final review) is in progress. See PROGRESS.md for the live task board and docs/BLUEPRINT.md for the full roadmap.
Full setup — external accounts, installs, running the pipeline, backend, and frontend locally — is maintained in docs/runbook.md. Short version:
git clone https://github.com/DevGurav/urbanheat-mumbai.git
cd urbanheat-mumbai
cp .env.example .env # fill in GEE_PROJECT_ID, GEMINI_API_KEY, SUPABASE_* (runbook.md §1)
uv sync --extra pipeline --group dev
uv run python -m data_pipeline.run --stage all # builds data/, models/ (runbook.md §8)
uv run uvicorn backend.main:app --reload # http://localhost:8000/docs
cd frontend && npm install && npm run dev # http://localhost:5173data_pipeline/ Earth Engine + OSM + weather extraction, ml/ (train, explain, HVI, scenario)
data/ Feature tables, rasters (large files gitignored — regenerate via pipeline)
backend/ FastAPI: routers/ agents/ rag/ auth.py store.py services.py
frontend/ Vite + React + TypeScript dashboard
supabase/ schema.sql — saved_scenarios table + RLS policies
notebooks/ Exploration and ML experiments
docs/ Architecture, decisions, methodology, runbook, devlog
.github/ CI + scheduled monitoring cron workflows
Dockerfile Backend image — built and pushed locally, not by CI (docs/runbook.md §4)
| Doc | Contents |
|---|---|
| BLUEPRINT.md | Master roadmap: phases, exit criteria, scope |
| conventions.md | Hard rules, Definition of Done, code conventions |
| architecture.md | Components, data flow, deployment topology |
| decisions/ | Architecture decision records — why each choice |
| data-dictionary.md | Every dataset and feature: source, units, licence |
| ml-methodology.md | Model design, validation strategy, metrics |
| agents.md | Agent roles, tools, prompts, guardrails |
| api-reference.md | Endpoint contracts |
| runbook.md | Setup, run, deploy, troubleshoot |
| devlog.md | Session-by-session engineering journal |
| references.md | Papers and datasets cited |
Code: MIT (see LICENSE). Data sources retain their own licences — Landsat/SRTM are
public domain (USGS/NASA), Sentinel-2 and ESA WorldCover are CC BY 4.0, WorldPop is
CC BY 4.0, OpenStreetMap is ODbL. Attribution details in
docs/data-dictionary.md.
Author: Devendra Gurav (@DevGurav)