About this research
This research forecasts when and where fog forms over rivers (“river fog”), using weather data and machine learning. This page is a pilot publication of the forecast for each daytime hour, up to 3 days ahead, at the river-camera sites in the Tadami River and Aizu area.
The content on this page is computed output from a research-stage forecast model; it is not a forecast from the Japan Meteorological Agency.
Scale of the data
〜2026/11/01
Method
- 1ObservationRiver-camera images from the Ministry of Land, Infrastructure, Transport and Tourism are collected hourly.
- 2ClassificationImages are classified by a multimodal AI (Google Gemini) for the presence of river fog. Manually checked/corrected records are also used.
- 3TrainingA gradient-boosted decision tree (LightGBM) learns the relationship between river fog and weather variables (temperature, humidity, wind, etc.) at the same time.
- 4ForecastingWeather forecast values are used to compute the probability of river fog for each site and hour. The model is retrained daily on the latest data. For the rest of the current day (up to 8 hours ahead), a separate model recomputes the forecast using the hourly camera observations (each site's latest fog status and the fraction of foggy sites overall) as additional input.
The 3-level forecast scale
| Likelihood: High | Hours where the forecast probability is at or above the decision threshold. The threshold is generally chosen to maximize the F1 score on validation data using past forecast values (the current threshold's precision is about 24%). |
| Likelihood: Medium | Hours where the probability is below the threshold but at or above 50% of it. |
| Likelihood: Low | All other hours. |
| (blank) | Hours before sunrise or after sunset. Camera images cannot be classified then, so these hours are excluded from the forecast. |
Forecast accuracy check
Evaluated on the most recent period not used for training (2026-10-11 – 2026-11-01), using the actual forecast values as input (daytime only).
Reference: forecast by climatology (past occurrence rate by site/hour) 0.23
River fog is rare — it occurs in only about 2% of daytime hours — so current forecast accuracy is limited. The model is continually improved against the benchmark of beating the climatology forecast.
Notes on use
- This page is a pilot publication for research purposes; its content or availability may change or stop without notice.
- Forecasts may contain errors. If using them to decide on travel, photography, sightseeing, etc., check official sources such as the Japan Meteorological Agency and use your own judgment.
- The “current conditions” field is an AI-classified observation from camera images and may contain misclassifications.
Data sources
- Weather forecast / past values: Open-Meteo (CC BY 4.0)
- Past weather observations: Japan Meteorological Agency
- River camera images: MLIT “River Disaster Prevention Information”
- Map: GSI Japan (GSI tiles, pale map)
Research staff can use the research site via “Staff login” at the top right.