Crop Fire Smoke Attribution for Lahore Smog Using Satellite Fire Data and Wind Back-Trajectories

The problem: everyone argues about smog, few measure its sources daily

Every October and November, Lahore and much of Punjab disappear under smog. Schools close, hospitals see more breathing problems, and the public debate goes in circles: crop residue burning, traffic, brick kilns, industry, weather, or fires across the border. All play a part, and the mix changes day to day with the wind.

Policy responses depend on which sources matter on which days. Crop burning bans, kiln closures and traffic restrictions all have costs. A transparent, daily estimate of how much smog is linked to upwind fires, built only from open data and published methods, would make the debate more honest and the response better targeted.

What already exists and the gap

• NASA FIRMS publishes near real time satellite fire detections worldwide, including fire radiative power, a measure of fire intensity.
• Air quality platforms such as IQAir and government monitoring networks report PM2.5 levels, which tell you how bad the air is, not where it came from.
• NOAA's HYSPLIT model computes air mass trajectories and is free to use for research.
• Research studies have estimated fire contributions to pollution in Delhi for specific seasons.

The gap is a daily, repeatable attribution for Lahore that combines these sources, shows its working and its uncertainty, and is published where anyone can check it.

Inspiration

Cusworth and colleagues' paper "Quantifying the influence of agricultural fires in northwest India on urban air pollution in Delhi, India" (Environmental Research Letters, 2018) combined satellite fire data, atmospheric transport and surface measurements to estimate the contribution of crop fires to Delhi's particulate pollution. Your project brings a simplified, data-driven version of that approach to Lahore and runs it every day.

What you build

• The fire pipeline downloads daily fire detections across both Punjabs and surrounding regions.
• The trajectory step computes back-trajectories from Lahore at several heights for each day and records which grid cells the air passed over, and when.
• The exposure index combines fire intensity along the trajectories, weighted by distance and time.
• The model explains daily PM2.5 with the fire index, weather (wind speed, boundary layer height, humidity) and seasonal terms, then estimates the share linked to fires.
• The dashboard shows today's trajectories, fires, PM2.5 and the estimated fire share with its range.

The technically interesting part

• Trajectories. Run HYSPLIT, or compute simpler trajectories yourself from ERA5 wind fields. Compare both for a sample of days and explain the difference.
• Exposure weighting. Smoke from a fire 40 kilometres away six hours ago matters more than one 300 kilometres away two days ago. Design and test the weighting, and show how sensitive results are to it.
• Confounding by weather. Calm, cold, stable days trap all pollution, local and remote. If you ignore weather, you will over-credit fires. Include weather terms and test whether the fire effect survives.
• Satellite limits. Haze, cloud and satellite overpass times mean some fires are missed. Note this as a source of underestimation.
• Uncertainty. Report the fire share as a range from bootstrap or model uncertainty, never as a single exact percentage.

How to evaluate it

• Model fit: out-of-sample error for daily PM2.5 on held-out seasons, with and without the fire index.
• Plausibility: whether the estimated fire share peaks during the known burning weeks and falls after them.
• Sensitivity: how much the attribution moves with different trajectory heights and weighting choices.
• Missing days: how many days lacked PM2.5 or fire data, and how the dashboard handles them.
• Literature check: whether your fire share range is broadly consistent with published studies for the region, and explanations where it is not.

Suggested plan

• Weeks 1 to 4: collect several seasons of fire, PM2.5 and weather data, explore daily patterns.
• Weeks 5 to 9: trajectories with HYSPLIT, exposure index, sensitivity tests.
• Weeks 10 to 14: regression model, attribution, validation.
• Second semester: daily automation, dashboard, comparison with literature, a second city if time allows, and the write-up.

Path to a product

Environmental agencies, newsrooms, research groups and climate-focused NGOs all need credible source information during smog season. An open daily attribution service with a methods paper behind it is a strong public good and a serious research output. Atmospheric data science and climate analytics roles value exactly this mix of satellite data, modelling and communication.

Scope warning

Do not attempt full chemical transport modelling; that is a doctoral project. Keep the attribution statistical and clearly labelled as an estimate. Avoid blaming named countries or groups in your summaries; report what the data shows, with its uncertainty, and let readers draw conclusions. Start with two past seasons before automating anything daily.

Learn what you need on Sefism

• Data engineering roadmap: reliable daily downloads and backfills.
• Data science roadmap: regression, confounding and uncertainty.
• Python Data Science Handbook: arrays and grids for wind and fire data.
• From Data to Viz: choosing charts for daily time series and maps.
• Our World in Data: background on air pollution and health for the write-up.

Get help with this idea

The trap is a confident percentage that ignores weather. Tauseef can help you design the sensitivity tests, plan the two semesters and present a contested topic carefully. 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