AfterPass: what burns after the satellite passes
Satellites count Punjab’s farm fires at 1:30 PM, but most are lit after 3. AfterPass finds the missing burns in Sentinel-2 imagery from AWS Open Data, and forecasts when the smoke reaches Delhi.
Every autumn, Delhi's air gets worse as paddy stubble burns in Punjab and Haryana, and every day the news reports a number: how many farm fires the satellites counted. That number comes from NASA's polar-orbiting satellites, which cross Punjab at about 1:30 PM. A fire is counted only if it is burning at that moment.
An iFOREST study using geostationary data, which looks every 15 minutes, found that most large farm fires in Punjab in 2024 and 2025 started after 3 PM. Those fires are out long before the next pass. AfterPass, built for the WeMakeDevs × AWS Environmental Hacks, asks a different question: not "what was burning at 1:30?", but "what has burned?"
The idea: look for the scar, not the flame
A burned paddy field turns black and stays black for days. Sentinel-2 photographs Punjab every two to five days at about 11 AM, in bands where char looks nothing like straw or soil. So AfterPass:
- Finds burn scars. For each clear Sentinel-2 image, a cropland pixel is flagged when it looks like char (low Normalized Burn Ratio, high mid-infrared burn index, dark in the near infrared) and its burn ratio dropped sharply since the last clear look. Flagged pixels become field-sized scars of half a hectare or more.
- Matches them with official detections. A scar is "in official counts" if a FIRMS detection lies within one sensor pixel of it (375 m for VIIRS, 1 km for MODIS), at the right time. Otherwise it is "not in official counts".
- Follows the smoke. Each scar releases model smoke between 3 and 7 PM on the days it could have burned. The smoke rides the 100 m wind from Open-Meteo to Delhi, where a Smoke Arrival Index scores each neighbourhood hour by hour for the next 48 hours.
- Answers questions in English and Hindi. An assistant built with Strands Agents answers only from the published data and moves the map to show what it is talking about.
Results: the 2025 season in three districts
For Sangrur, Bathinda and Firozpur, 1 October to 30 November 2025:
- 423,996 hectares burned, and 303,189 hectares (71%) had no matching official detection.
- Of 4,377 official detections, 4,330 (98.9%) were made between noon and 4 PM, around the afternoon pass. Our own data shows the blind spot.
- The village at the top of the "send help here first" list, Chaunda in Sangrur, had 850.6 hectares of burning not in official counts.
- Over 59 days, the daily Delhi-wide smoke index and PM2.5 from 40 CPCB and DPCC monitors had a Pearson correlation of 0.56 (Spearman 0.63).
Honest limits
- Burned area is a lower bound. Clouds leave gaps; a field tilled under before the next clear image is missed; so are fields under half a hectare.
- We don't know when a field was lit, only between which two images it burned.
- Calibration against FIRMS is a weak check. A VIIRS pixel is 375 m wide, so many pixels next to a detection never burned, and the measured recall (32%) understates what we find. The real test is a visual check of 200 random samples, which is built but not yet labelled, so precision is not yet known.
- The smoke correlation is partly weather. A shallow night-time mixing layer raises both the index and PM2.5, so the correlation shows the two move together, not how much PM2.5 the stubble smoke adds. The index is relative and is never shown as µg/m³.
The stack
Everything is on AWS, in the Sydney region:
- AWS Open Data for the heavy inputs: Sentinel-2 L2A cloud-optimized GeoTIFFs (through Element 84's Earth Search), ESA WorldCover for the cropland mask, and Overture Maps for village and Delhi locality names (queried with DuckDB straight from S3).
- Lambda and Step Functions for the daily pipeline: fetch FIRMS, find scenes, detect burns per district in a Map state, match and aggregate, then smoke and imagery in parallel, publish, and alert. EventBridge Scheduler runs it at 7:12 PM IST, after FIRMS has processed the afternoon pass, with a smoke-only refresh at 7:12 AM.
- S3 and CloudFront serve the published JSON and imagery to the app.
- Amazon Bedrock, through Strands Agents, runs the assistant on a Lambda Function URL, with a DynamoDB rate limit.
- SNS emails Delhi smoke alerts by zone; Amazon Location Service draws the basemap with India's political view; Amplify Hosting serves the React app.
The frontend is React with deck.gl on MapLibre: extruded H3 columns for burned area, FIRMS detections as cyan points, smoke as animated trails, and a react-globe.gl hero.
What fought back
- Region and model access. The plan was
us-west-2, next to the Sentinel-2 data, with Claude on Bedrock. The account could create resources only in Sydney and had no Claude quota. The data still streams fromus-west-2, and the assistant runs onopenai.gpt-oss-120bthrough Bedrock's OpenAI-compatible endpoint. Switching to Claude is a one-line setting once quota exists.
- Grounding an open model. Every answer is checked mechanically: each number in it must appear in that turn's tool results. All 10 scripted questions, five in Hindi, passed.
- Reading COGs reliably. Long runs of HTTP range reads occasionally came back truncated. Reads are aligned to the files' internal blocks, run in parallel, and retried.
- IST is UTC+5:30. Days in India start at half past the UTC hour, which broke an assumption that model hours start on UTC hours.
- Thousands of smoke trails in a browser. Unix-second timestamps lose precision in the GPU's 32-bit floats, so trail times are stored relative to each run. Trails are capped at 15,000, with a point every two hours, to keep each 48-hour run near 4 MB.
- Not overstating anything. The hardest rule was the one about numbers: everything on screen and in the assistant's answers comes from a pipeline output file, with its
generated_at.
Try it
- Method and limits:
docs/METHOD.mdin the repository.
AfterPass shows villages and districts, never individual farmers. Its purpose is to send crop-residue machines and awareness work to where uncounted burning concentrates, and to give Delhi a warning before the smoke arrives.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article