Each dated map uses the latest available landscape, calendar, weather and vegetation features with the saved XGBoost model to rank cells across Japan, based on patterns associated with past satellite fire detections.
1. Model building and evaluation
- Data preparation. A cell-day is one grid cell on one UTC date. Training combines fire cell-days, which have a recorded MODIS fire detection, with sampled non-fire cell-days, which have none.
- Held-out evaluation. Each input combination is tested on one year while the model trains on the other eight. Repeat for each year from 2017 to 2025.
- Input selection. We compared different combinations of inputs and chose landscape, calendar date, temperature, wind, vegetation greenness (NDVI) and vegetation moisture (NDMI).
- Final model. The selected model is trained on the full 2017–2025 training sample and saved for dashboard updates.
Loading evaluation calculations…
How results are evaluated
- Top-5% capture: the share of fire cell-days found in the highest-scoring 5% of the weighted test sample. This is not 5% of the map area.
- ROC AUC: how well scores rank fire cell-days above non-fire cell-days. A score of 1 means perfect separation; 0.5 is random ranking.
- Average precision (AP): how concentrated fire cell-days are near the top of the ranking, averaged as more of them are included. Higher values mean a better ranking.
- Random-ranking references for the same nine years: capture 13.1%, ROC AUC 0.500 and average precision 0.060. These reflect the weighted test samples and calculation method.
2. Daily updates
- Daily check. Updates are checked at 09:15 JST (00:15 UTC), after the UTC assessment date changes. If sources are not ready, the last published assessment remains available.
- Latest inputs. Weather uses the latest available daily observations and MODIS the latest completed 8-day composite. Fixed layers retain their source years. Open the information button beside Assessment date for source dates.
- Cell gaps. Missing inputs use the training median and a missingness flag. Cells with no selected weather or vegetation values remain unscored.
- Scoring. The saved model scores and ranks eligible cells without retraining. Automated checks verify the data and scores before the dashboard updates.