KilnTrace

• 

  KilnTrace: AI finds brick kilns from space, evidence locked on the Polygon blockchain

• 

  Live dashboard: 889 kilns in Kasur District, coloured by compliance state

• 

  Click any kiln: satellite chip, type, confidence, and in-browser on-chain verification

• 

  Verifier: real data passes all 5 checks; one edited field fails at the Merkle root

• 

  Precision–recall on the held-out test set: 0.917 mAP50 for kiln detection

• 

  Training label check: random SentinelKilnDB tiles with oriented boxes drawn

Track 4: Climate Data & Environmental Monitoring

Inspiration

Brick kilns are a major source of Punjab's winter smog. The law requires ~10,000 kilns to use cleaner zigzag technology, but checking is manual, slow and disputed. AI can map kilns from space, but nobody has a record everyone trusts as unaltered.

What it does

• AI finds every kiln in free Sentinel-2 satellite images and flags non-zigzag (non-compliant) ones.
• Tracks change since 2023: possible conversions, new and closed kilns.
• Locks the evidence on Polygon: all records go into a Merkle tree, and one transaction stores its root, the AI model's hash and the code version.
• Anyone can verify any record in the browser. Change one field and verification fails.

Kasur District: 730 kilns found, 387 non-zigzag flagged, 86 possible conversions, 889 records on-chain.

How we built it

YOLO11-OBB trained on SentinelKilnDB (0.917 mAP50, above the dataset paper's 0.875) · Sentinel-2 via Planetary Computer · Solidity + OpenZeppelin on Polygon Amoy · Python verifier · MapLibre + ethers.js dashboard on GitHub Pages.

Challenges

• Our first run missed most zigzag kilns. The training images used a different pixel grid, and December smog hid kiln colour. Matching the grid and using clear Feb–Mar imagery raised zigzag detections from 16 to 169.
• Python, Solidity and JavaScript must hash identical bytes, so records contain no decimals.

What we learned

Matching how training data was prepared matters as much as the model. A blockchain doesn't make data true; it makes it impossible to change silently.

What's next

Field-check app (contract already supports it), thermal data for smog-ban violations, all of Punjab.

Built With

• maplibre
• openzeppelin
• polygon
• python
• pytorch
• solidity
• ultralytics
• web3.py