Unlicensed fireworks and pyrotechnic manufacturing site detection
Clandestine pyrotechnic factories hide in farm sheds and residential plots, but thermal infrared and sub-metre optical imagery expose the physical and thermal signatures that licensing authorities never see on paper.
Sensors
- Maxar WorldView-3: Panchromatic resolution of 0.31 m and multispectral at 1.24 m enables identification of individual blast berms, dispersed storage pods and surface staining from oxidiser drying. On-demand tasking with same-day collection possible over most of Asia.
- Airbus Pléiades Neo: 0.30 m panchromatic, 1.2 m multispectral, with a daily revisit capacity over any point on Earth. Stereo and tri-stereo modes allow 3-D reconstruction of berm height, which correlates with the quantity of explosive material a site is designed to contain.
- Landsat 8/9 TIRS: Thermal Infrared Sensor records land surface temperature in two bands (10.6–11.2 µm and 11.5–12.5 µm) at 100 m native resolution (resampled to 30 m). Sufficient to flag persistent anomalous warmth from large drying floors or mixing operations against a cool agricultural background. Free archive back to 2013.
- ECOSTRESS (ISS): NASA's ECOsystem Spaceborne Thermal Radiometer Experiment on Space Station acquires land surface temperature at approximately 70 m × 70 m pixel size with variable overpass timing, which means it can capture thermal state at different times of day than Landsat. Useful for confirming diurnal thermal persistence rather than a single nocturnal anomaly.
Why clandestine pyrotechnic sites look the way they do
Pyrotechnic manufacturing is unusually dangerous even when done correctly. Oxidisers such as potassium nitrate and potassium perchlorate are mixed with fuels and metal powders in small batches precisely because a single ignition event can detonate adjacent stock. Licensed facilities are therefore required by national and international standards to physically separate each stage of production, store intermediate product in dispersed magazines, and construct blast-resistant barriers between buildings. These requirements are not optional engineering choices; they are the physical geometry of the risk.
Unlicensed operators face the same physical risk but no regulatory inspection. They replicate the dispersed layout informally, because the alternative is a single catastrophic loss. The result is a characteristic spatial signature: a cluster of small structures spaced 15 to 40 metres apart, often with low earthen or brick berms between them, set within or adjacent to an agricultural building that provides cover from casual observation. Blast walls made from compacted earth are cheaper than concrete and leave a distinctive low-relief ridge visible in sub-metre stereo imagery.
What a floating roof gives away
The thermal signature of pyrotechnic manufacturing is a product of two operations: drying and mixing. Oxidiser compounds are often purchased or synthesised in solution and must be dried to a target moisture content before blending. Drying floors, even when covered by a thin corrugated roof to hide them from casual aerial observation, radiate measurable heat. At Landsat TIRS resolution, a 20-by-30-metre drying apron operating at 10 to 15 degrees Celsius above ambient will produce a detectable anomaly against a rice-paddy background, which typically sits within 2 to 3 degrees of the regional mean during the growing season.
ECOSTRESS adds a useful dimension here. Because the International Space Station precesses across local solar time over roughly 60 days, ECOSTRESS captures the same site at different times of day across a multi-month observation window. A thermal anomaly that persists from mid-morning through late afternoon is consistent with an active drying or curing process rather than a transient agricultural burn or cooking fire. That diurnal persistence is a meaningful discriminator. It is not conclusive on its own, but it narrows the candidate list considerably before a commercial tasking order is placed.
Reading the optical signature at sub-metre resolution
Once a thermal candidate is identified, WorldView-3 or Pléiades Neo imagery resolves the physical layout. The features that matter are specific. Blast berms appear as low linear ridges casting shadows at oblique sun angles; their height can be estimated from shadow length and sun elevation angle, giving a rough proxy for the explosive mass the operator expects to contain. Dispersed outbuildings with no apparent agricultural function, separated from the main structure by irregular spacing inconsistent with storage efficiency, are a second indicator. Surface staining from oxidiser compounds, particularly potassium nitrate, can produce pale crystalline deposits visible in multispectral imagery.
A licensed facility will typically have a single formal access road, a perimeter fence of consistent construction, and a clear separation between the production zone and any residential structure. Clandestine sites often show the opposite: multiple informal tracks, mixed residential and production use within the same compound, and ad-hoc screening such as bamboo matting or tarpaulins over open areas. None of these features is individually diagnostic. The value is in the combination, scored against a signature library built from documented enforcement cases in India, the Philippines and Indonesia, where post-explosion investigations have produced detailed site photographs and floor plans that are publicly available through official inquiry reports.
Honest limits: what the sensors cannot resolve
Thermal detection at Landsat resolution fails for small operations. A single-room mixing operation inside a brick building with a tiled roof may produce no detectable surface temperature anomaly at 30-metre pixels. ECOSTRESS at 70 metres is worse in this respect. The thermal approach is calibrated to sites with active drying floors of at least 100 to 200 square metres, which corresponds to a medium-to-large clandestine operation rather than a household-scale one.
Cloud cover is a persistent problem across South and Southeast Asia during the monsoon season, which can interrupt optical and thermal collection for weeks at a time. Synthetic aperture radar can detect structural changes through cloud, but SAR at commercially available resolutions (Sentinel-1 at 5 by 20 metres in IW mode, or ICEYE and Capella at 0.5 to 1 metre in spotlight mode) does not reliably distinguish blast berms from agricultural earthworks without supporting optical context. Archive depth helps: Landsat's free record back to 2013 allows analysts to identify when a dispersed outbuilding cluster appeared, which is useful for enforcement chronology but not for real-time interdiction.
From pixel to prosecution file
The analytic workflow moves in three stages. First, a thermal screening pass over a defined area of interest uses Landsat TIRS time series to rank grid cells by anomaly frequency and magnitude. Cells that show persistent elevated temperature in more than 30 percent of cloud-free observations over a 12-month window are flagged for optical review. Second, a sub-metre optical collection is tasked against the flagged cells to assess physical layout against the signature library. Third, confirmed candidates are packaged as georeferenced site reports with annotated imagery, estimated site age from archive comparison, and a confidence tier.
Satellize applies this workflow under client licence, drawing on open thermal data and commercial optical tasking. The same multi-source approach that underpins the Kingdom of Tonga crop-estimation programme, combining open and commercial imagery in a single analytic pipeline, applies here, though the detection problem is obviously different. Enforcement agencies typically need output in formats compatible with their GIS environments; delivery as georeferenced GeoTIFF overlays and PDF site reports with coordinate tables is standard.
Scoping a monitoring programme
The practical starting point for a licensing authority or national police agency is a thermal screening pass over the districts with the highest historical incident rates. In India, for example, Tamil Nadu and West Bengal account for a disproportionate share of documented factory explosions; in the Philippines, Bulacan province has a long-documented concentration of unlicensed producers. A province-scale thermal screening using the full Landsat archive costs nothing in data fees and can be completed analytically in days.
Optical tasking is the cost driver. A single WorldView-3 or Pléiades Neo collection over a 25-square-kilometre area of interest runs to a few hundred US dollars at standard commercial rates, though pricing varies by provider and contract structure. The economics favour a tiered approach: screen broadly with free thermal data, task commercially only against confirmed thermal candidates. A programme covering a single high-risk district might require 10 to 20 commercial collections per year, with the remainder of the workflow running on open data.
Typical figures
| Optical resolution (commercial) | 0.30 m panchromatic (Pléiades Neo, WorldView-3); 1.2–1.24 m multispectral |
| Thermal resolution | 100 m native / 30 m resampled (Landsat 8/9 TIRS); ~70 m (ECOSTRESS) |
| Thermal bands | 10.6–11.2 µm and 11.5–12.5 µm (Landsat TIRS bands 10 and 11) |
| Revisit (optical tasking) | Daily or better over most of Asia (Pléiades Neo constellation); same-day possible (WorldView-3) |
| Revisit (thermal, free) | 16-day exact repeat (Landsat); variable local solar time over ~60-day cycle (ECOSTRESS) |
| Minimum detectable thermal anomaly | ~100–200 m² active drying floor at ≥10 °C above ambient (Landsat TIRS); smaller anomalies below detection threshold at this resolution |
| Archive depth | Landsat: free back to 2013 (TIRS); commercial optical: typically 3–5 years depending on provider |
| Cloud limitation | Optical and thermal collection blocked by cloud; SAR (Sentinel-1, ICEYE, Capella) provides structural change detection through cloud but with reduced feature discrimination |
| Delivery formats | Georeferenced GeoTIFF, KMZ/KML, shapefile, PDF annotated site report with coordinate tables |
| Latency (commercial tasking to delivery) | Typically 24–72 hours from collection to analysed output, depending on cloud and processing queue |
Analytics Satellize can run
| Thermal anomaly screening layer | Landsat TIRS time-series land surface temperature retrieval; anomaly frequency scoring over 12-month rolling window | GIS polygon layer ranking grid cells by anomaly persistence; updated quarterly or on request |
| ECOSTRESS diurnal persistence report | Multi-overpass LST comparison across local solar time to distinguish active thermal process from transient burn events | Per-candidate summary table with overpass times, temperature delta and persistence classification |
| Sub-metre site layout assessment | Object-based image analysis of WorldView-3 or Pléiades Neo imagery; feature extraction for berm geometry, outbuilding spacing and surface staining | Annotated imagery with labelled features, berm height estimate from shadow analysis, confidence tier (low / medium / high) |
| Site age and change chronology | Archive optical comparison (Landsat, Sentinel-2, Planet where licensed) to identify first appearance of dispersed outbuilding cluster | Timeline graphic and table for enforcement case file, showing construction sequence |
| Compound signature scoring | Multi-feature scoring against documented signature library derived from post-explosion investigation reports (India, Philippines, Indonesia) | Structured site report with scored feature checklist and overall risk classification |
| Area-wide candidate ranking for enforcement prioritisation | Combined thermal and optical screening across a defined administrative area; ranked candidate list by composite score | Priority-ranked site list in PDF and GIS format, suitable for field inspection scheduling |
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.