OSINT · thermal sources

Thermal evidence beyond a hotspot.

Support industrial thermal-source and fire investigations with persistent-source attribution, detection-aware inference, abstention, potential missed-source review and observation planning.

Workflow reference: THERMA-ID demo. Based on its published description; this workbench supplies inference primitives, not the demo's trained model or validation results.

Live thermal evidence and private AOI watches

Real NASA FIRMS detection centroids from four rolling sensor feeds are collected into a durable index every ten minutes by default. Live requests use a five-minute cache. Review observation dates separately from retrieval time: the source may be delayed. Points do not identify fire boundaries or prove industrial source classes. Missing detections are not observed absences.

Evidence area (west,south,east,north)

Thermal evidence

FIRMS detections and thermal imagery record observed heat, rather than automatically proving a source class.

Explore evidence →

Infrastructure context

Review OSM completeness, land cover and environmental evidence before specifying class priors.

Explore evidence →

Forecast and observation conditions

Use ECMWF first, then NOAA/NWS, alongside cloud, viewing geometry and sensor detectability.

Explore evidence →

Review NASA FIRMS, ECMWF forecasts and artifact footprints. A recurring AOI watch helps collect new evidence; the analysis below runs on your supplied observations and is not an automatic trained monitoring service.

Detection-aware thermal analysis

Define 2–65 class/phase states, a prior and transition matrix, and up to 366 regularly spaced observations. Supply each state's probability of a positive detection, including false positives and cloud/geometry effects. Observed absence is false; an unobserved pass is null. OSM completeness weakens the supplied prior toward a uniform distribution. The example is illustrative and uncalibrated. Add a WGS84 location record with geometry, bbox or longitude/latitude to export a footprint; simulated defaults to true. Optional candidate_passes accepts up to 64 uniquely named passes, each with source, steps_ahead (1–30 model steps) and detectability per state. To evaluate predictions, add an optional label to an observation with class, source and optional active (true/false). Labels never affect inference; the example labels are synthetic.

Model and evidence JSON

{
  "model_provenance": "Illustrative analyst model; replace with validated site and sensor parameters",
  "location": null,
  "simulated": true,
  "osm_completeness": 0.6,
  "abstention_threshold": 0.8,
  "states": [
    {
      "class": "industrial",
      "phase": "active",
      "active": true
    },
    {
      "class": "industrial",
      "phase": "idle",
      "active": false
    },
    {
      "class": "vegetation fire",
      "phase": "active",
      "active": true
    },
    {
      "class": "vegetation fire",
      "phase": "idle",
      "active": false
    }
  ],
  "prior": [
    0.35,
    0.35,
    0.15,
    0.15
  ],
  "transition": [
    [
      0.8,
      0.18,
      0.01,
      0.01
    ],
    [
      0.2,
      0.78,
      0.01,
      0.01
    ],
    [
      0.01,
      0.01,
      0.7,
      0.28
    ],
    [
      0.01,
      0.01,
      0.08,
      0.9
    ]
  ],
  "candidate_passes": [
    {
      "id": "clear-pass",
      "source": "Illustrative clear-sky sensor model",
      "steps_ahead": 1,
      "detectability": [
        0.85,
        0.05,
        0.9,
        0.03
      ]
    },
    {
      "id": "cloudy-pass",
      "source": "Illustrative cloudy sensor model",
      "steps_ahead": 1,
      "detectability": [
        0.1,
        0.01,
        0.15,
        0.01
      ]
    }
  ],
  "observations": [
    {
      "datetime": "2026-09-27T00:00:00Z",
      "detected": true,
      "detectability": [
        0.85,
        0.05,
        0.9,
        0.03
      ],
      "source": "Example clear-sky observation",
      "label": {
        "class": "industrial",
        "active": true,
        "source": "Synthetic example label; not verified ground truth"
      }
    },
    {
      "datetime": "2026-09-28T00:00:00Z",
      "detected": false,
      "detectability": [
        0.1,
        0.01,
        0.15,
        0.01
      ],
      "source": "Example cloudy observed pass",
      "label": {
        "class": "industrial",
        "active": true,
        "source": "Synthetic example label; not verified ground truth"
      }
    },
    {
      "datetime": "2026-09-29T00:00:00Z",
      "detected": null,
      "detectability": [
        0.85,
        0.05,
        0.9,
        0.03
      ],
      "source": "Example missing observation"
    }
  ]
}

Saved monitoring runs

Save the last analyzed model and its results for repeatable review. Optionally link a real thermal watch to snapshot every indexed NASA detection in its AOI between the model's first and last observation timestamps. Evidence is context only and does not change your supplied observations or detectability. Runs are saved when you request them; recurring analysis is not scheduled.

Optional real thermal watch ID

Keep run and watch IDs private: possession grants access. Saved runs have no public listing and are not isolated by user accounts. Each packet is limited to 8 MiB and an evidence snapshot to 10,000 detections; requests that exceed a limit fail explicitly.

Private saved run ID