Next 48-Hour PM2.5 Forecasts for South Asian Cities from Open Sensors and Weather Data
Each winter, smog in Lahore, Delhi, Dhaka and Kathmandu closes schools and harms health, and decisions are made with little warning. Build a model that forecasts PM2.5 for the next 48 hours at city sensor locations, using open sensor readings, weather forecasts and fire detections, and compare it against the global CAMS forecast. Present it as a simple outlook a school or clinic can use to plan.
Intermediate
A small team, or one strong student willing to learn something new.1 semester · 2 students
7 to ship
4 optional extrasSuggested stack
What you should ship
- A data pipeline joining hourly PM2.5 readings from OpenAQ or official monitors with weather data and satellite fire counts
- Data quality checks for stuck, missing and implausible sensor readings
- Baselines: persistence, daily cycle and the CAMS global forecast at the same location
- A gradient boosted model forecasting PM2.5 at 6, 12, 24 and 48 hours
- Prediction intervals and the probability of crossing health category thresholds
- Evaluation over at least one full winter season held out from training
- A public-facing outlook page in English and a local language
If you have time left
- Forecasts for locations between sensors using spatial interpolation
- Adding low-cost sensor data with correction for humidity
- An alert email for schools when a threshold crossing is likely
- Explaining forecasts with feature attributions (wind, fires, inversion)
The problem: smog decisions made in the morning
From November to January, PM2.5 levels in cities across the Indo-Gangetic plain often reach hazardous levels. Schools close at short notice, outdoor workers keep working, and hospitals see more breathing problems. Most people check current readings, which tell them what is already happening. A reliable 24 to 48 hour outlook would let schools, clinics, event organisers and families plan.
What already exists and the gap
- CAMS, the Copernicus Atmosphere Monitoring Service, publishes global air quality forecasts, at coarse resolution.
- IQAir and similar platforms show current readings and some forecasts.
- Government environment agencies publish monitor readings and sometimes advisories.
Global forecasts often miss local peaks, and current readings do not help planning. The gap is a city-level forecast trained on local sensor data, evaluated honestly against the global baseline.
Inspiration
OpenAQ, a nonprofit that aggregates air quality data from governments and other sources worldwide, makes the training data possible. Studies using machine learning to correct or downscale CAMS forecasts with local measurements have shown gains over raw global output; your project tests that idea for South Asian winter smog.
What you build
- Schools and clinics see a 48-hour outlook for their area with the chance of reaching each health category.
- Journalists and residents get a plain explanation of what drives the forecast this week.
- Researchers get a clean, reproducible dataset and benchmark.
The technically interesting part
- Sensor data quality. Monitors go offline, get stuck or report impossible values. Build checks and log what you removed.
- The right features. Wind speed and direction, boundary layer height, humidity, temperature inversions, upwind fire counts and day of week all matter. Use forecast weather, not observed weather, for future hours, or you are cheating.
- Beating the baseline. Compare against persistence and CAMS. If you only beat persistence on easy days, say so.
- Communicating uncertainty. Probabilities of crossing a threshold are more useful than a single number.
How to evaluate it
- Error by lead time (6 to 48 hours) against persistence and CAMS.
- Threshold skill: hit rate and false alarm rate for crossing the hazardous category.
- Calibration of the predicted probabilities.
- Held-out winter: train on earlier seasons, test on the most recent full winter.
Suggested plan
- Weeks 1 to 3: collect data for one or two cities and build quality checks.
- Weeks 4 to 6: baselines and CAMS comparison.
- Weeks 7 to 10: model, intervals and threshold probabilities.
- Weeks 11 to 13: evaluation and the outlook page.
- Weeks 14 to 15: report and demo.
Path to a product
Schools, hospital groups, factories with outdoor workers and air purifier companies all want better local forecasts. A forecasting API for South Asian cities, built on open data and validated in public, is a credible offering. It is also a strong data science portfolio piece.
Scope warning
Do not present your forecast as an official advisory. Credit data sources and follow their licences. Keep to one or two cities at first.
Learn what you need on Sefism
- Data science roadmap: forecasting, intervals and evaluation.
- Data engineering roadmap: the ingestion pipeline and quality checks.
- Machine learning roadmap: gradient boosted models.
- DuckDB documentation: fast local queries on hourly data.
- Streamlit documentation: the outlook page.
- Our World in Data: background on air pollution and health.
Get help with this idea
The usual mistake is using observed weather for future hours, which makes results look far better than they are. Tauseef can help you set up the evaluation correctly. If you are new here, create a free account first.
Ideas and guidance, not finished projects
These are project ideas and scoping guidance, published free for students to use as a starting point. I do not build, write, or sell final-year projects, and I do not complete coursework for anyone. Take an idea, make it yours, and build it.
Not sure this one fits you?
I help students pick an idea that matches their skills and their deadline, then scope it down until it is finishable. Guidance only, never done for you.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Sessions are free for approved Sefism members, and a few slots open each week.
Follow along
New writing, resources and project ideas land here first.