🌊 Lumina

176 lakes, tracked live. Ten years of satellite history, backfilled to 2015. New imagery every 5 days at 10 m per pixel. A colorblind-safe health scale. $0 per lake, MIT licensed.

Lumina turns free Sentinel-2 imagery into a live, public intelligence platform for urban lakes — measuring water loss, algae blooms, and recovery for every lake in a city, every few days. It finally reads the photographs.

🛰️ The Quiet Crisis

Every lake on Earth is photographed every five days. Nobody reads the photographs.

That single observation is the heart of Lumina. Cities bleed their lakes away — encroached, silted, blooming with algae — and the official record is a PDF table somebody drew once a year. The photograph of the problem exists. It's just unread.

The data is captured; comprehension is the bottleneck.


🏙️ Vanishing lakes	Encroachment, sewage, drought; Bengaluru lost dozens of ~200 lakes in one generation
🧪 Unscalable monitoring	Boats, lab samples, funding — most lakes are never tested; by the time it's visible from shore, it's often irreversible
🗄️ Locked data	PDFs and silos, invisible to citizens, activists, planners
🌡️ Planetary stakes	Lakes are carbon sinks, habitats, flood buffers, groundwater recharge

🔬 The Pixel That Tells the Truth

Lumina closes the loop — computed server-side on Google Earth Engine.

  Sentinel-2 passes overhead (every ~5 days)
                    │
                    ▼
   ┌───────────────────────────────────────┐
   │  📡 Pull imagery automatically          │
   │  💧 NDWI + adaptive Otsu threshold      │
   │     → exact water surface area          │
   │  🦠 NDVI within the water mask          │
   │     → algae blooms & weed cover         │
   │  💯 Health score 0–100                  │
   │     trend · algae · volatility          │
   └───────────────────────────────────────┘
                    │
                    ▼
  🗺️ Live map · 🏆 Leaderboards · 🎞️ Timelapses · 📊 Charts

Answers in one glance:

🔴 Which lakes are in crisis now (ranked health)
📉 Which lakes shrink (10-year trends)
🦠 Which bloom (NDVI proxy)
🆚 What changed since 2017 (drag-to-compare)

Colorblind-safe by design — teal, not green; distinct luminance at every stop; survives grayscale.

🌅 A Decade of History

Ten years backfilled on day one, so trends surface in weeks — while restoration is still possible.


🛰️ Sentinel-2 archive	COPERNICUS/S2_SR_HARMONIZED, 10 m/pixel
☁️ Lake-level cloud filter	Rejected only if the lake is cloudy — rescues monsoon scenes
💧 NDWI + Otsu adaptive threshold	Self-calibrates per scene — no fragile fixed cutoff
🦠 NDVI algae estimate	Within the water mask
💯 Health Score 0–100	Trend + algae + seasonal volatility
⚖️ Relatable units	"≈35 football fields lost" — citizens feel the number

🌍 From Orbit to Action

Remote sensing stops being a science project when it becomes civic infrastructure.

Earth Forward pillar	Mechanism	Outcome
🌊 Water conservation	Per-lake area tracked every 5 days, decade-deep	Shrinkage detected in weeks
🌡️ Climate resilience	Quantify failing flood buffers/recharge	Cities see failure before the flood
🦆 Ecosystems	Automatic algae-bloom detection + health scoring	Pollution surfaces without dead fish
🏘️ Communities	Free dashboard, relatable units, shareable cards	One trusted picture for all
⚖️ Equity	$0/lake, open source, any city deployable	Monitoring as public infrastructure

UN SDGs: 🎯 SDG 6 · SDG 11 · SDG 13 · SDG 15

🧭 The System, Instrumented

Three independently deployable components, one city-agnostic config:

Component	Stack	Role
`pipeline/`	Python 3.12 · Google Earth Engine · SQLAlchemy	Ingest, classify, score, render timelapse
`backend/`	FastAPI · PostgreSQL	REST: lakes, observations, leaderboards
`frontend/`	Next.js 16 · React 19 · Mapbox GL · Tailwind 4 · Recharts	Public dashboard

🔄 Self-updating forever — GitHub Actions cron re-ingests every 5 days; zero manual work.

🏙️ Any city, one config — drop in city.config.json + a lake-boundary GeoJSON.

🏔️ The Climb

☁️ Monsoon clouds — judge cloudiness over each lake polygon, not the tile
📐 Adaptive Otsu thresholding — per-scene NDWI cutoff, not a planet-wide guess
💾 Petabyte-scale, zero downloads — server-side GEE compute; only numbers return
🎨 A scale everyone can read — colorblind-safe by design
💬 Numbers that feel — football fields, not hectares

📖 What We Learned

We came in knowing web development; we left knowing remote sensing. NDWI/NDVI, Otsu thresholding, Sentinel-2 band math, GEE's server-side model, seasonal time-series, timelapse rendering — and the truth that good environmental tech is 20% science, 80% communication.

🔭 After the Horizon

🌧️ Rainfall correlation — drought vs. encroachment
🚨 Early-warning alerts on sharp score drops
🌐 Multi-city federation
🧪 Ground-truth integration with lab and field data
🛰️ Sentinel-1 SAR through monsoon clouds

🎞️ The Record, on Screen

Lumina map dashboard — every lake in Bengaluru colored by live health score

The city at a glance — every lake as a polygon colored by its health score. Lumina leaderboard — lakes ranked by what the satellites see

The leaderboard — Most Shrunk, Most Recovered, Most Polluted, Most Volatile. Lumina lake detail page — Sentinel-2 timelapse and health score

Every lake's dashboard — playable 10-year timelapse and relatable shoreline stats. Then vs now — draggable comparison of 2017 and 2025 satellite frames

Then vs now — draggable comparison of 2017 and 2025 satellite frames

Then vs now — drag the slider. The change is undeniable.

🥇 The Case for Lumina

Originality: Orbital Earth observation as a civic product: backfilled decade, explained trends, then-vs-now, per-lake timelapses
Earth Forward: Ecosystem monitoring + conservation + community resilience, implemented fully
Completion: Live dashboard, real lakes, real data, leaderboards, charts, timelapses, compare slider, self-updating cron
Learning: Web dev → remote sensing, spectral indices, adaptive thresholding, cloud geospatial compute
Design: Calm dark map-first UI, colorblind-safe scale, relatable units, social-ready cards
Technology: GEE server-side pixel math + Otsu + NDWI/NDVI + volatility scoring + 5-day ingestion + GIF/MP4 timelapse + FastAPI + Next.js