Agricultural crop-residue burning and acute PM2.5 episode attribution
Post-harvest burning in South Asia and sub-Saharan Africa produces PM2.5 spikes that overwhelm ground monitors. VIIRS fire radiative power, MODIS aerosol optical depth, and HYSPLIT back-trajectories can attribute specific burning events to specific downwind populations.
Sensors
- VIIRS VNP14 Active Fire (Suomi-NPP / NOAA-20 / NOAA-21): 375 m pixel resolution fire detection with fire radiative power (FRP) in megawatts per pixel. Twice-daily overpass at any given latitude. FRP is a quantitative proxy for fuel consumption rate and therefore emission intensity, not merely fire presence. Minimum detectable fire size is roughly 0.1 ha under favourable geometry, though small smouldering fires are routinely missed.
- MODIS MAIAC AOD (Terra / Aqua): Multi-Angle Implementation of Atmospheric Correction retrieves aerosol optical depth at 1 km resolution, improving on the standard 10 km Dark Target product over heterogeneous land. Twice-daily combined Terra/Aqua coverage. Cloud contamination and bright surfaces (desert, snow) produce data gaps; over-ocean retrievals are more reliable than over arid land.
- TROPOMI CO columns (Sentinel-5P): Carbon monoxide total column at 5.5 km x 7 km pixel (post-2019 reprocessing). CO is a conservative fire tracer: it persists for weeks in the free troposphere, making it useful for tracking aged smoke far downwind of the source. Daily global coverage. Sensitivity degrades in the boundary layer, where CO concentrations matter most for human exposure.
- NOAA HYSPLIT trajectory model: Hybrid Single-Particle Lagrangian Integrated Trajectory model, run operationally by NOAA ARL. Back-trajectories compute the most probable air-mass origin for a receptor point over 24, 48 or 72 hours. Ensemble runs with perturbed meteorological fields give uncertainty bounds. Trajectory error grows with time; beyond 72 hours, attribution becomes indicative rather than definitive.
- FIRMS near-real-time fire alerts (NASA): Fire Information for Resource Management System aggregates VIIRS and MODIS detections and delivers alerts with latency as low as 3 hours from satellite overpass. Useful for operational episode onset detection, though the latency floor is set by overpass timing, not data processing.
Why stubble fires are a different problem from background pollution
Agricultural residue burning is episodic and geographically concentrated. In the Indo-Gangetic Plain, rice-straw burning in Punjab and Haryana runs for roughly three to four weeks each October and November, immediately after the kharif harvest and before the rabi planting window. Farmers burn because it is fast and cheap; the alternative, mechanical incorporation, costs money they often do not have. The result is a pulse of PM2.5 that can push Delhi's air quality index above 400 in 48 hours, a concentration roughly 16 times the WHO 24-hour guideline of 15 µg/m³.
The policy and legal challenge is attribution. Urban monitors record the spike, but a city government cannot act on a source 300 km upwind without evidence linking that source to the receptor. Ground-based source apportionment (chemical mass balance, receptor modelling) is expensive, slow, and requires a dense monitoring network that most affected cities in South Asia and sub-Saharan Africa do not have. Satellite-derived attribution fills that gap, provided the analyst is honest about what the data can and cannot resolve.
What fire radiative power actually measures, and what it does not
VIIRS VNP14 reports fire radiative power in watts per square metre, integrated over the 375 m pixel. FRP correlates with fuel consumption rate through a well-established empirical relationship: roughly 0.368 g of dry matter consumed per megajoule of radiant energy released, per the Wooster et al. formulation that underpins most operational fire emissions work. Multiply by emission factors for PM2.5 (which vary by crop type, moisture content and combustion phase) and you get an estimated emission flux. For rice straw, published emission factors cluster around 6 to 9 g of PM2.5 per kilogram of dry matter burned, though the range across studies is wide.
The honest limits matter here. VIIRS misses fires smaller than roughly 0.1 ha, misses smouldering combustion (low temperature, low FRP), and produces no detection under thick cloud. In the pre-monsoon burning season in sub-Saharan Africa, cloud cover is lower and detection rates are higher. In the post-monsoon Punjab window, cloud is less of a problem, but haze from earlier fires can attenuate the thermal signal. FRP-based emission estimates carry uncertainties of 30 to 50 percent even under good conditions.
Connecting the fire to the city: HYSPLIT back-trajectory mechanics
A back-trajectory answers a specific question: given that air is arriving at receptor point R at time T, where was that air parcel 24, 48 or 72 hours earlier? NOAA HYSPLIT, run against NCEP GDAS or ERA5 reanalysis winds, produces a set of trajectory lines. Overlaying those lines on the VIIRS fire detections from the preceding days identifies which burning areas the air mass traversed.
This is not a dispersion model. HYSPLIT does not compute concentration at the receptor; it traces the most probable path of an air parcel. To get concentrations, analysts use HYSPLIT in forward-trajectory mode with emission inputs, or couple it to a chemical transport model such as GEOS-Chem or WRF-Chem. The simpler back-trajectory approach is sufficient for attribution (did air from burning region X reach city Y?) but not for quantifying the PM2.5 contribution in µg/m³. That distinction matters when the output is used in litigation or regulatory proceedings.
TROPOMI CO columns add a second line of evidence. A CO enhancement over a city that is spatially and temporally coherent with upwind fire detections and consistent with the back-trajectory path is strong corroborating evidence of fire origin. CO from traffic and industry has a different spatial signature: diffuse, persistent, tied to road networks rather than appearing as a plume.
Separating agricultural fire from the urban source mix
Cities in the Indo-Gangetic Plain and in West African urban corridors have multiple simultaneous PM2.5 sources: diesel vehicles, brick kilns, biomass cookstoves, industrial stacks. An acute episode is rarely pure agricultural smoke. The attribution task is to partition the observed AOD or PM2.5 spike into its contributing fractions.
Three observational signals help. First, fire-episode AOD tends to be spatially coherent over large areas (hundreds of kilometres) rather than localised near point sources. Second, TROPOMI CO and, where available, TROPOMI formaldehyde (HCHO) can distinguish combustion-dominated from photochemical-dominated aerosol loading. Third, the timing of the AOD spike relative to the fire detection peak, lagged by the transport time implied by the back-trajectory, provides a temporal fingerprint. None of these signals is individually conclusive. Used together, with honest uncertainty bounds, they support defensible attribution at the episode level, not the individual-fire level.
Latency, coverage and the operational window that matters
For a health-warning application, the useful window is 24 to 72 hours before a population is exposed. FIRMS delivers VIIRS fire alerts within 3 hours of overpass. HYSPLIT forward trajectories can be run in minutes once fire locations and FRP values are in hand. MAIAC AOD from the same day's Terra and Aqua passes is typically available within 6 to 12 hours. That pipeline is fast enough to support next-day air quality advisories.
The binding constraint is usually not data latency but meteorological uncertainty. Wind forecast errors at 48 hours can displace a predicted plume by 100 km or more, which at urban scales is the difference between a health alert and a false alarm. Ensemble trajectory runs, using multiple meteorological realisations, give a probability distribution over affected areas rather than a single deterministic footprint. Communicating that uncertainty to public health officials, without losing the operational utility of the warning, is as much a design problem as a data problem.
Satellize runs this attribution pipeline on open constellations for clients who need it embedded in existing health-surveillance workflows. The crop-estimation programme in the Kingdom of Tonga uses a related fire and vegetation monitoring stack, which gives the team direct operational experience with VIIRS-based agricultural signal extraction in data-sparse environments.
What the numbers can support, and what they cannot
A well-executed satellite attribution study can establish, with reasonable confidence, that a specific multi-day burning episode in a defined agricultural region contributed materially to an observed PM2.5 spike at a downwind receptor city. It can rank source regions by estimated emission contribution. It can identify the days on which transport conditions were most favourable for impact.
It cannot, with current open-constellation data, resolve individual farm-level burning events at sub-hectare scale in near-real time. It cannot replace ground-based chemical speciation for source fingerprinting when legal liability is at stake. And it cannot fully account for the boundary-layer dynamics that determine whether smoke aloft actually reaches surface-level breathing zones. Those limits are not reasons to avoid the analysis; they are reasons to frame the outputs honestly and pair them with whatever ground monitoring exists.
Typical figures
| Fire detection spatial resolution | 375 m (VIIRS VNP14); 1 km (MODIS MOD14/MYD14) |
| AOD spatial resolution | 1 km (MODIS MAIAC); 3.5 km x 5.5 km (TROPOMI aerosol index) |
| CO column resolution | 5.5 km x 7 km (TROPOMI, post-2019 reprocessing) |
| Revisit frequency | Twice daily per sensor; VIIRS + MODIS combined gives up to 4 daytime passes at mid-latitudes |
| FIRMS alert latency | As low as 3 hours from satellite overpass to public alert |
| HYSPLIT trajectory duration | Operationally run at 24, 48 and 72 hours; uncertainty grows substantially beyond 48 hours |
| Minimum detectable fire (VIIRS) | Approximately 0.1 ha under favourable geometry and clear sky; smouldering fires routinely missed |
| FRP-based emission uncertainty | 30 to 50 percent under good conditions; higher under partial cloud or thick haze |
| AOD data gaps | Cloud cover, sun glint and bright surfaces (desert, snow) cause systematic gaps; gap frequency highest in monsoon season |
| Archive depth | VIIRS from 2012 (Suomi-NPP); MODIS from 2000 (Terra) and 2002 (Aqua); TROPOMI from 2018 |
Analytics Satellize can run
| Daily active fire map with FRP-weighted emission index | VIIRS VNP14 fire detection aggregated by administrative boundary; FRP summed and converted to estimated PM2.5 emission flux using published crop-type emission factors | GIS layer (GeoJSON or shapefile) updated daily during burning season, with per-district emission ranking |
| Episode onset alert | Threshold exceedance on 3-day cumulative FRP within a defined source region, triggering automated notification | JSON alert feed with source region, cumulative FRP, estimated onset date, and link to trajectory forecast |
| Back-trajectory attribution report | HYSPLIT ensemble back-trajectories (24/48/72 h) from receptor city, overlaid on VIIRS fire detections; CO column enhancement used as corroborating tracer | PDF episode report with trajectory maps, fire detection overlay, CO anomaly panel, and written attribution confidence statement |
| Seasonal burning calendar by source region | Multi-year VIIRS archive aggregated by dekad (10-day period) to characterise typical burning onset, peak and cessation dates by district | Tabular dataset and chart pack for health-system pre-positioning and advance warning planning |
| AOD anomaly time series for receptor cities | MODIS MAIAC AOD extracted for urban polygons, baseline-corrected against same-period multi-year median, flagged for episode days exceeding 2 standard deviations | CSV time series with episode flags, suitable for correlation with hospital admissions data held by the client |
| Source-region contribution ranking | Forward HYSPLIT dispersion runs from multiple source districts, weighted by VIIRS FRP, to estimate relative PM2.5 contribution at receptor; presented as probability-weighted ranking, not point estimates | Ranked table with confidence bands, accompanied by methodology note on model uncertainty |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.