Open-source tools to preserve, reconstruct and study wildfire events.
FireViewer started during the wildfire in Die, France, the town where I grew up and where part of my family still lives.
At the time I was mainly trying to understand what was happening from scattered official information, maps, images, videos and geographic data.
The project grew from there.
What started as a way to make sense of one event gradually became an attempt to answer a more difficult question:
Can we preserve a wildfire well enough that it can be reopened, checked and studied later instead of being reduced to a final map?
Today FireViewer combines evidence collection, computer vision, satellite observations, deterministic geographic processing, daily fire-state reconstruction and reproducible OpenUSD environments.
Its main rule is simple:
observed ≠ reconstructed ≠ simulated ≠ predicted
A useful reconstruction does not become an observation just because it looks convincing.
A model output is not automatically evidence.
And when the available information is not sufficient, unknown or abstain
is a better result than invented precision.
FireViewer is not an emergency alert service, an official wildfire source, an incident-command system or a certified fire-spread predictor.
flowchart TB
SOURCES["Official · public · authorised sources"] --> EVIDENCE["Evidence<br/>time · provenance · rights"]
EVIDENCE --> GEO["Deterministic geographic hypotheses"]
EVIDENCE --> SAT["Dated satellite observations"]
GEO --> ASSESS["Multimodal assessment<br/>accept · reject · abstain"]
ASSESS --> PART4["Part.4 3.3<br/>daily reconstruction"]
SAT --> PART4
SEED["Private dated<br/>initial affected area"] --> PART4
PART4 --> FROZEN["Frozen state<br/>lineage · uncertainty"]
FROZEN --> REVIEW["Review · corrections"]
REVIEW --> INCIDENT["Versioned incident archive<br/>2D · 3D · later study"]
MAP["Part.1 Map Builder<br/>measured OpenUSD context"] --> INCIDENT
FROZEN --> EVAL["Isolated evaluation"]
REF["Held-out references"] --> EVAL
The separation is intentional.
Evaluation references do not feed reconstruction.
Measured maps do not become reconstructed wildfire state.
Synthetic scenes do not become evidence of real events.
FireViewer is still a research MVP, but a substantial technical foundation already exists:
- bounded evidence acquisition and provenance;
- image and video processing;
- deterministic geographic hypotheses;
- satellite evidence from several source families;
- structured multimodal assessment with explicit abstention;
- Part.4 3.3 daily reconstruction;
- versioned probability, provenance and spatial artifacts;
- reproducible measured-map production;
- OpenUSD and web-view spatial packages;
- synthetic-data tooling kept separate from real evidence;
- public models, datasets and measured-map resources.
The complete real-data path is not yet qualified as an unattended production service, and the current Part.4 profile remains uncalibrated.
That distinction matters: code existing and something being proven to work reliably in the real world are not the same thing.
The current reconstruction line starts from a private, dated estimate of the initially affected area.
That initial contour does not mean the whole area is active.
Later admissible observations contribute to daily versions of:
affectedactiveobservable- uncertainty
The system keeps probability state, source lineage and revisions instead of silently rebuilding history from the latest geometry.
A correction creates another revision. It does not erase the previous one.
Historical reconstruction also respects time: a satellite product available today is not automatically considered information that was available at the historical date being reconstructed.
More details:
The Map Builder is a separate part of FireViewer.
It creates measured geographic environments and portable OpenUSD packages from versioned geographic inputs.
These environments provide spatial context in which an incident can later be explored or studied.
They are not generated wildfire boundaries.
The environmental asset library is also still relatively small. Improving its quality, diversity and reproducibility is one of the next important steps for FireViewer.
FireViewer publishes research models, datasets, measured maps and reproducibility material through:
https://huggingface.co/fireviewer
Resources can have different states: active, research, restricted, legacy or superseded.
Limitations and failed assumptions are useful information too; not every experiment needs to be presented as a successful production model.
FireViewer is currently mostly maintained by me, and there is no large community behind it yet.
That is fine.
Useful contributions do not need to be large.
Finding a broken OpenUSD package, checking a geographic assumption, testing a dataset, improving accessibility, reproducing a bug, correcting documentation or showing that one of my approaches is wrong can all help.
See:
For now, technical governance is deliberately simple and maintainer-led.
A French non-profit association is being created to give FireViewer a proper administrative and financial structure.
The association is meant to help the project survive and grow — not to create an artificial layer of technical authority above the people actually working on it.
The governance should reflect the project that actually exists, not imitate the governance of a much larger project we do not have yet.
FireViewer has so far been developed with limited financial resources and a large amount of personal time.
Infrastructure credits, compute, storage, reusable data, technical help, scientific review, OpenUSD/geospatial expertise, grants and other forms of support can all be useful.
Support does not buy influence over evidence, uncertainty or technical results.
For research, infrastructure, collaboration, rights, provenance or security: