Citizen Science Game for Verifying Brick Kiln Detections from Satellite Images

The problem: smog sources nobody has fully mapped

Every winter, cities across the Indo-Gangetic Plain, including Lahore, Delhi and Dhaka, sit under heavy smog. Brick kilns are one of the significant local sources, and older designs burn far less cleanly than newer zig-zag kilns. Governments have pushed for kilns to convert to cleaner designs, and some have ordered closures during the worst weeks. Enforcement depends on knowing where every kiln is and which design it uses, and that map is hard to keep current: kilns open, close and convert, often in rural areas.

Machine learning on satellite images can find many kilns automatically, but models make mistakes, especially in new regions. Checking thousands of detections by hand is slow for a small team, and an unchecked map is easy for anyone to dispute.

What already exists and the gap

• Research models and datasets exist. A Stanford-led team detected brick kilns across Bangladesh from satellite imagery, and a 2025 dataset in Scientific Data maps kilns in Pakistan's part of the Indo-Gangetic Plain.
• Zooniverse runs hundreds of citizen science projects in which volunteers classify images, and has passed a billion classifications.
• Environmental agencies rely on field inspections and complaints.

The gap is a focused, engaging way to verify and correct model detections at scale, with statistics that turn volunteer answers into labels a regulator or researcher could trust, and an open map as the result.

Inspiration

"Scalable deep learning to identify brick kilns and aid regulatory capacity" (Lee, Brooks and colleagues, PNAS, 2021) located kilns across Bangladesh with high accuracy, distinguished the banned older design from cleaner ones by shape, and linked kiln emissions to air quality in Dhaka. The authors framed the work as support for regulators. Your project adds the missing human step: a game in which many people verify the model's findings, combined with the classic Dawid and Skene method (1979) for estimating each annotator's error rate and the true label from many noisy answers.

What you build

• Volunteers, such as students, play short rounds on a phone or laptop. Each round shows image tiles with a candidate kiln; players confirm, reject or mark the design. Points depend on agreement with consensus and on gold-standard tasks.
• Researchers and NGOs see a live map of verified kilns, each with a confidence level, and download it as open data.
• The model (stretch goal) improves by retraining on verified labels.

The technically interesting part

• Preparing tasks. Run an existing detector on open imagery for one district, crop tiles around each candidate, and include some random tiles without kilns so players cannot simply confirm everything.
• Measuring players. Mix in gold-standard tiles with expert labels. Estimate each player's accuracy, and use it to weight their answers.
• Consensus. Implement Dawid-Skene with expectation maximisation, compare it with simple majority voting, and decide how many answers each tile needs before it is considered settled.
• Game design that rewards care. Points for speed would ruin the data. Reward accuracy, streaks of correct gold answers and finding missed kilns.
• Difficult tiles. Some tiles stay disputed even after many answers. Send those to experts, and use them to learn which situations confuse both players and the model.
• Spatial data. Store candidates and verified kilns with PostGIS, deduplicate detections that overlap, and export GeoJSON.

How to evaluate it

• Label quality: accuracy of consensus labels against expert labels on a held-out sample of at least 200 tiles, compared with majority voting and with the model alone.
• Efficiency: average number of player answers needed per settled tile.
• Engagement: rounds played, return rate and drop-off points during the playtest.
• Model improvement (stretch): detector accuracy before and after retraining on verified labels.

Suggested plan

• Weeks 1 to 3: read the kiln papers, obtain imagery and an existing model or dataset, choose one district.
• Weeks 4 to 6: candidate pipeline, tiles and gold-standard set.
• Weeks 7 to 10: game, scoring and backend.
• Weeks 11 to 13: consensus engine, map and open data export.
• Weeks 14 to 15: playtest with volunteers, evaluation, write-up and demo.

Path to a product

Environmental agencies, air quality researchers and NGOs need up-to-date maps of pollution sources, and the same game and consensus engine can verify other features, such as solar panels, flood damage or informal settlements. A grant-funded open platform, or a verification service for research groups, are both realistic. The project is a distinctive portfolio piece that joins games, computer vision and data quality.

Scope warning

Do not train a new detection model from scratch; use published models or datasets and focus on verification and consensus. Keep to one district at first. Respect the licence terms of the imagery you use, and present the map as research data, not as evidence against any individual business.

Learn what you need on Sefism

• Game development roadmap: short rounds, scoring and motivation.
• Machine learning roadmap: running a detector and measuring label quality.
• Data science roadmap: the statistics behind combining many noisy answers.
• torchvision Documentation: image models and preprocessing.
• PostgreSQL Documentation: storing and querying the kiln map with PostGIS.
• FastAPI Tutorial: User Guide: the backend for tasks and answers.

Get help with this idea

The risk is a fun game that produces labels nobody can trust, or a year lost on imagery. Tauseef can help you choose the district and data source, design the gold-standard tasks and plan the consensus evaluation. If you are new here, create a free account first.

• Discuss this idea with a mentor
• Send a query to Tauseef
• Book a 1:1 session