Presentation Information
[O12-P83]Analysis of the Visibility of Mt. Fuji and Its Characteristic Clouds
*Shunta Umino1, *Eito Usami1, *Shiori Hoshino1 (1. Tokyo Metropolitan Tachikawa Senior High School)
Keywords:
Machine Learning,Mt.Fuji,YOLOv9,Clouds,Windprofiler
[Background and Objective]
The Astronomy and Meteorology Club at our school has conducted meteorological observations using a Stevenson screen for about 80 years, measuring temperature, pressure, visibility, and Mt. Fuji observations twice daily. Visual observations of Mt. Fuji’s visibility increased in missing data after 1995 but were resumed in 2018 following Taguchi (2019). However, frequent missing observations led to the consideration of automation.
In 2020, an automated imaging system using a Raspberry Pi to control a DSLR camera was developed, and a large number of images have since been accumulated. Mt. Fuji, an isolated peak, often forms characteristic clouds such as cap clouds, lenticular clouds, apron clouds, and banner clouds, which have long been used for traditional weather forecasting.
Kusaka et al. (2025) clarified relationships between upper-level wind and temperature for cap, lenticular, and banner clouds using mesoscale analysis data, but relationships with apron clouds, pressure patterns, and surrounding weather remained unclear. Therefore, this study aims to analyze approximately 200,000 images collected over five years to count cloud occurrences and examine their relationships with meteorological elements using accessible data.
Additionally, because analyzing visibility requires manual inspection of large numbers of images, we aimed to automate classification using machine learning. Since visual observations often included ambiguous cases, a three-class dataset (“visible,” “partially visible,” and “not visible”) was constructed.
[Methods]
① Analysis of Clouds and Meteorological Elements
Approximately 200,000 images taken between October 2020 and December 2024 were visually inspected to identify four types of clouds: cap, lenticular, apron, and banner clouds. Occurrences were aggregated by month and time of day.
For each event, pressure patterns, saturation levels at the summit from 12 hours before to 1 hour after, and wind direction and speed at different altitudes were obtained. Wind profiler diagrams were created to analyze wind conditions. Additionally, precipitation at the Kawaguchiko AMeDAS station after cloud formation was examined and compared with average precipitation probability.
② Automatic Classification of Visibility
A dataset of 990 images was collected and classified into three categories—“visible,” “partially visible,” and “not visible”—based on defined criteria, and annotated. A machine learning model was developed using transfer learning with the YOLOv9 pre-trained model (gelan-c.pt).
[Results and Discussion]
① Clouds and Meteorological Elements
Monthly analysis showed that banner clouds were most frequent in winter, and occurrences peaked in the morning (6–9 AM). Under a typical winter pressure pattern (west-high, east-low), about 80% of clouds were banner clouds.
Cap and lenticular clouds showed increasing saturation near the summit during formation, with only cap clouds reaching full saturation. This likely reflects their formation heights, with cap clouds forming near the summit and lenticular clouds at higher, unsaturated levels.
Wind analysis indicated west-southwesterly winds for cap and lenticular clouds (stronger for lenticular clouds), and strong west-northwesterly winds for banner clouds. Precipitation probability was highest during lenticular cloud events and lowest during banner cloud events. Apron clouds may have been misclassified due to difficulty in identifying their position.
② Automatic Visibility Classification
The model achieved a high F1 score of 0.93 when classifying 996 images. It was integrated into an automated system for image capture, classification, and recording, and linked to a web database to display updated records.
Analysis of 2025 data showed that Mt. Fuji is more visible in winter and less visible in summer. The “partially visible” category accounted for only 1–6% throughout the year, with no significant seasonal variation.
[Future Work]
A new automated system with a 300 mm telephoto lens will be introduced to improve classification accuracy. Broader AMeDAS data will be used to enhance predictions of precipitation and surrounding weather.
Further analysis will examine seasonal trends including cases where Mt. Fuji is 50–90% visible within the “partially visible” category. Additionally, visual classification of 2025 images will be conducted to compare with automated results and further evaluate model performance.
The Astronomy and Meteorology Club at our school has conducted meteorological observations using a Stevenson screen for about 80 years, measuring temperature, pressure, visibility, and Mt. Fuji observations twice daily. Visual observations of Mt. Fuji’s visibility increased in missing data after 1995 but were resumed in 2018 following Taguchi (2019). However, frequent missing observations led to the consideration of automation.
In 2020, an automated imaging system using a Raspberry Pi to control a DSLR camera was developed, and a large number of images have since been accumulated. Mt. Fuji, an isolated peak, often forms characteristic clouds such as cap clouds, lenticular clouds, apron clouds, and banner clouds, which have long been used for traditional weather forecasting.
Kusaka et al. (2025) clarified relationships between upper-level wind and temperature for cap, lenticular, and banner clouds using mesoscale analysis data, but relationships with apron clouds, pressure patterns, and surrounding weather remained unclear. Therefore, this study aims to analyze approximately 200,000 images collected over five years to count cloud occurrences and examine their relationships with meteorological elements using accessible data.
Additionally, because analyzing visibility requires manual inspection of large numbers of images, we aimed to automate classification using machine learning. Since visual observations often included ambiguous cases, a three-class dataset (“visible,” “partially visible,” and “not visible”) was constructed.
[Methods]
① Analysis of Clouds and Meteorological Elements
Approximately 200,000 images taken between October 2020 and December 2024 were visually inspected to identify four types of clouds: cap, lenticular, apron, and banner clouds. Occurrences were aggregated by month and time of day.
For each event, pressure patterns, saturation levels at the summit from 12 hours before to 1 hour after, and wind direction and speed at different altitudes were obtained. Wind profiler diagrams were created to analyze wind conditions. Additionally, precipitation at the Kawaguchiko AMeDAS station after cloud formation was examined and compared with average precipitation probability.
② Automatic Classification of Visibility
A dataset of 990 images was collected and classified into three categories—“visible,” “partially visible,” and “not visible”—based on defined criteria, and annotated. A machine learning model was developed using transfer learning with the YOLOv9 pre-trained model (gelan-c.pt).
[Results and Discussion]
① Clouds and Meteorological Elements
Monthly analysis showed that banner clouds were most frequent in winter, and occurrences peaked in the morning (6–9 AM). Under a typical winter pressure pattern (west-high, east-low), about 80% of clouds were banner clouds.
Cap and lenticular clouds showed increasing saturation near the summit during formation, with only cap clouds reaching full saturation. This likely reflects their formation heights, with cap clouds forming near the summit and lenticular clouds at higher, unsaturated levels.
Wind analysis indicated west-southwesterly winds for cap and lenticular clouds (stronger for lenticular clouds), and strong west-northwesterly winds for banner clouds. Precipitation probability was highest during lenticular cloud events and lowest during banner cloud events. Apron clouds may have been misclassified due to difficulty in identifying their position.
② Automatic Visibility Classification
The model achieved a high F1 score of 0.93 when classifying 996 images. It was integrated into an automated system for image capture, classification, and recording, and linked to a web database to display updated records.
Analysis of 2025 data showed that Mt. Fuji is more visible in winter and less visible in summer. The “partially visible” category accounted for only 1–6% throughout the year, with no significant seasonal variation.
[Future Work]
A new automated system with a 300 mm telephoto lens will be introduced to improve classification accuracy. Broader AMeDAS data will be used to enhance predictions of precipitation and surrounding weather.
Further analysis will examine seasonal trends including cases where Mt. Fuji is 50–90% visible within the “partially visible” category. Additionally, visual classification of 2025 images will be conducted to compare with automated results and further evaluate model performance.