tinychaseLab · a transformer over FIRMS detection sequences
Fire, Flare or Volcano?
Every hot pixel a satellite sees becomes a FIRMS detection, and TerraPulse clusters those into thermal events over eight weeks, all of them typed unclassified. A wildfire, a gas flare and a lava lake are all hot, but they are not hot in the same rhythm. Can a small transformer, reading only the sequence of detections, tell them apart with no hand labels at all?
The world, typed
Every cluster in the eight-week window, typed by the model that needs no dossier (summary numbers and how often the same spot lit up), with its odds corrected back to how rare flares and volcanoes are in the manifest. Each cell is two degrees; amber where a quarter or more of its clusters are called flare, blue where a quarter or more are called volcano, size is how many clusters it holds. The savanna burning season is the grey mass across Africa, South America and northern Australia.
Does order matter?
Four models, three seeds, scored on held-out weak labels from map cells never trained on. Macro-F1 over the three classes: a model that only guessed the majority would score 0.33. The two MLPs see the same information as the transformer, but as averages; the gap between them and it is what the sequence adds.
The one honest test: FIRMS's own labels
228 clusters from an August 2020 ingest pilot in California carry the FIRMS type flag. The model never saw 2020, never saw a label from FIRMS, and here is what it calls each type.
Basins nobody labelled
The flare label came only from US oil-and-gas facilities. These are the world's other flaring basins, where no cluster was labelled anything, and what the model calls them.
The volcano label, audited
Volcano meant "within 3 km of a Wikidata volcano", which also catches a grass fire on a dormant cone. What the sequence model says about each of the most-labelled sites.
Three rhythms
The most confidently typed validation example of each class. Each dot is a detection: height is fire radiative power, and night-time passes (local 18:00 to 06:00) are filled. Suomi NPP passes near 01:30 and 13:30 local, so a site can light up at most twice a day; the question is which of those passes it lights up on, how brightly, and for how long.
How the sequence is tokenised
A cluster becomes at most 32 positions. Position 0 is a site token from the manifest: detection-count bin, how many other clusters sat within 1 km in the window, duration, extent, mean and peak power. Positions 1 to 31 are the detections in time order, evenly thinned when there are more. Each detection carries seven tokens whose embeddings are summed: local solar hour (24 bins), gap since the previous detection (same pass, under 1 h, 3 h, 6 h, 12 h, 18 h, 30 h, 48 h, longer), power in half-decade bins, brightness in 10 K bins, FIRMS confidence, distance from the centroid, and pixels lit in the same pass. Everything is a bin, so the model reads shape, not magnitude. Pretraining hides 15% of positions and asks the encoder to guess their power, hour and gap; the typing head is then fitted to weak labels: within 1.5 km of an EPA oil-and-gas extraction facility is flare, within 3 km of a volcano is volcano, and far from both and outside any known flaring basin is fire.