How to read the map
The default colours compare all served cells across Japan. The optional within-zone view ranks cells only against others in the same coarse model-defined geographic zone. Zone names are navigation and validation strata, not administrative boundaries. A top tier is a relative priority group, not a probability or warning.
Colour shows the susceptibility tier. Opacity separately indicates how much of an approximately 1 km sampling cell is natural-burnable. The underlying latitude/longitude lattice is not an equal-area grid.
What this map is for
Use the map alongside local records and field knowledge:
- Find a location by coordinates and inspect its landscape rank.
- Compare places and identify questions for local investigation.
- Download the scores for a transparent comparison with other evidence.
The model also ranks cells with no historical detection. That capability is useful for asking research questions, but the comparisons below do not establish that it selects better prevention sites than simple baselines. Fuel treatments, firebreaks and infrastructure siting require evidence about local fuels, spread, exposure, costs and intervention effects.
It is not suitable for same-day decisions. Do not use it for dispatch, staffing a shift, evacuation, closure decisions, or any warning function: it contains no weather, no ignition timing, and no calibrated probability, and the evidence below does not support a short-term target list. Low rank does not establish safety; unscored land has not been assessed.
For current conditions and official information, use the JMA forest-fire prevention portal and information issued by local authorities.
Scientific boundary
The current v4 surface is static and uses a legacy FIRMS target that could not exclude every persistent non-vegetation hotspot. It therefore represents susceptibility to historical satellite-detected fire activity, not confirmed wildfire incidence, current weather, ignition timing, impact, or calibrated occurrence probability.
A standard-source, presumed-vegetation-fire replacement is under prospective review and is not shown on the public map.
Evidence and baseline comparisons
In historical testing, the landscape model separates detected from undetected cells with a mean held-region ROC-AUC of 0.744, ahead of a coordinates-only control in all eight validation strata. This measures discrimination across geographic model zones; it does not establish practical prevention benefit.
| Compared cells | Landscape | Coordinates only |
|---|---|---|
| All served cells | 47.5% | 57.3% |
| Previously undetected cells | 31.5% | 37.1% |
The coordinate baseline captures more detected cells at this budget. These are retrospective comparisons using contemporary landscape layers, not untouched future tests or calibrated probabilities.
Forward testing on data collected after the model was frozen is a much sterner test, and it has barely begun. One qualifying date has been evaluated: predictions archived on 24 August 2026, evaluated for 25 August through 1 September. A count of dates alone cannot establish validation or justify promotion. In this window, 33 served cells had detections, distributed among disjoint within-model-zone bands (different from national ranks):
- Top 0.3% tier
- 1,339 cells — 0 of 33 detections
- Top 0.3–2% band
- 7,568 cells — 0 of 33 detections
- Top 2–10% band
- 35,613 cells — 18 of 33 (55%)
- Remaining cells
- 400,646 cells — 15 of 33 (45%)
Across the whole ranking the model ordered cells better than a trivial “this cell has recorded a detection before” baseline (ROC-AUC 0.90 versus 0.78). But that baseline concentrated its hits far more tightly, achieving roughly 2.8 times the average precision. This does not by itself prove superiority at every possible investigation budget.
Recalculation at a fixed national 5% cell budget captures 27.3% of those outcome cells with the landscape rank, versus 58.7% expected from the recurrence baseline with random selection among ties. The paired 50-km block interval for the difference is wide (−83.8 to +43.6 percentage points). At a 5% geographic-area budget, capture is 24.2% versus 58.7%. These intervals condition on the fixed selections in this one window.
Of 54 source detection rows, 40 assigned to the served lattice and reduced to 33 outcome cells; 14 rows could not be assigned to served cells. The reported full coverage concerns the assigned cells. Repeat detections are not separate incidents, and these outcomes are not confirmed wildfires.
This single inspected window supports no claim of operational value. A frozen future protocol will compare equal budgets across a full seasonal period and report spatial uncertainty. Passing software and source checks is separate from demonstrating useful predictions.
Current map contract
- Coverage
- 445,166 fraction-gated natural-burnable cells across Japan
- Model
- Fixed 24-feature XGBoost landscape susceptibility surface
- Primary comparison
- National percentile rank
- Optional comparison
- Within-model-zone priority
- Model vintage
- v4, frozen August 2026
- Detection archive
- 2017–2025; static landscape layers have separate vintages
- Presentation updated
- 7 September 2026; scores unchanged
- Time meaning
- Static historical susceptibility; no daily forecast
- Public output
- Read-only map, documented result table, and metadata