Yonaguni Visibility Analysis — 4,800 Days of Data from Japan's Westernmost Island

24.5m average, 99.4% of days above 15m — Yonaguni is Japan's clearest ocean. But AI can't predict it at all (R²≈0.05). 4,826 days of data and the mystery of this unique sea.

What emerged from this massive dataset, however, was deeply unexpected: Yonaguni is the single most difficult site for AI-based visibility prediction across all sites in our system. The general AI fails catastrophically here, and even a dedicated site-specific AI barely outperforms random chance. This article explores what 4,800 days of data reveal about Yonaguni's ocean, and why artificial intelligence cannot "understand" this sea.

Seasonal Visibility Patterns: Where Okinawa's Rules Do Not Apply

The monthly visibility profile of Yonaguni defies the typical Okinawa pattern. At most Okinawan dive sites, visibility peaks during summer when weather is calm and rainfall is relatively low. Yonaguni, however, does not follow this rule. The Kuroshio Current's main flow passes directly south of the island, and its micro-scale meandering patterns exert an outsized influence on daily visibility conditions. As a result, visibility at Yonaguni is governed more by moment-to-moment current dynamics than by seasonal weather cycles.

The island's reputation as a "big animal" destination centers on its winter season, when vast schools of hammerhead sharks patrol the surrounding waters. Yet winter also brings harsh sea conditions, with strong northeasterly winds limiting accessible dive points. Visibility records during winter months are biased toward sheltered southern and western points, adding another layer of complexity to the data.

The Catastrophic AI Failure: AI Predictions Less Accurate Than Simple Average

Our project deploys a general AI model trained on data from 30+ dive sites across Japan. For most sites, this model provides reasonable predictions, with the best-performing site (Izu Oceanic Park) achieving AI accuracy 82%. At Yonaguni, however, the general AI produces predictions that are less accurate than a simple average. This means the model's predictions are worse than simply predicting the site's historical average every single day. The model is not merely failing to learn Yonaguni's patterns; it is actively producing harmful predictions by applying patterns learned from other sites that do not apply here.

In response, we trained a dedicated site-specific AI using only Yonaguni's 4,826 observations. The result was AI accuracy 2%, meaning the model explains just 2.3% of visibility variance. For comparison, Izu Oceanic Park's site-specific AI achieves AI accuracy 82%, and even Kushimoto, a site also strongly influenced by the Kuroshio, reaches AI accuracy 42%. Despite having the largest dataset of any site, Yonaguni's visibility remains essentially unpredictable with our current set of weather, marine, and satellite variables.

Comparison with Other Okinawa Sites: Same Prefecture, Different Ocean

Comparing Yonaguni with other Okinawan dive sites highlights just how anomalous it is. Ishigaki Island, located 127 km to the east, achieves good prediction accuracy with our general AI. The Kerama Islands near Naha also show predictable seasonal patterns. At these sites, standard meteorological variables such as wind speed, rainfall, and wave height show a strong relationship with visibility.

At Yonaguni, these same predictors carry almost no information. The fundamental difference lies in geography and oceanography. While Ishigaki sits within a sheltered reef lagoon environment, Yonaguni is an isolated oceanic island at the boundary between the Pacific Ocean and the Philippine Sea. The seafloor drops precipitously to hundreds of meters around the island, creating complex current interactions, localized upwelling zones, and rapid water mass exchanges that no standard weather API can capture.

Long-term Yearly Trends

The yearly dataset spans from 2007 to 2026, providing one of the longest continuous diving visibility records in Japan. This multi-decadal perspective allows us to assess whether long-term climate trends, ocean warming, or changes in Kuroshio behavior have affected Yonaguni's underwater conditions. The inter-annual variability is substantial, further underscoring the difficulty of prediction.

Large-scale climate modes such as El Nino and La Nina may contribute to year-to-year fluctuations, potentially by shifting the Kuroshio's path or altering the volume of water flowing through the Taiwan Strait. However, establishing clear causal links requires more sophisticated analysis and additional oceanographic data sources that we aim to incorporate in future model iterations.

Hammerhead Season and Diving Conditions

Yonaguni's primary draw for international divers is the winter hammerhead shark aggregation (December through March). Schools of scalloped hammerheads numbering from dozens to occasionally hundreds pass through the waters around the island, making it one of the premier hammerhead destinations in the world, alongside the Galapagos, Cocos Island, and Malpelo.

Winter diving conditions, however, are demanding. Strong northeasterly monsoon winds frequently restrict access to southern and western dive sites. Drift diving in powerful currents is the norm, and advanced open water certification with significant current experience is strongly recommended. Visibility is highly variable on a day-to-day basis. When the Kuroshio flows directly past the island, visibility can exceed 30 m, offering spectacular encounters with hammerheads in crystal-clear water. When the current shifts, visibility can drop dramatically within hours. This rapid, unpredictable variability is precisely what makes Yonaguni so resistant to AI modeling.

Practical Diving Advice

Hammerhead Season (December-March)

This is the peak season for shark encounters. Water temperature ranges from 22 to 24 degrees Celsius, comfortable in a 5 mm wetsuit. Expect strong winds and potentially rough boat rides to dive points. Seasickness prevention is essential. Current experience and drift diving skills are mandatory; this is not a destination for beginners.

Underwater Monument (Year-round)

The mysterious "Yonaguni Submarine Ruins" off the island's southwest coast feature stepped stone formations at depths of 5 to 25 m. Whether natural geological formations or remnants of an ancient civilization remains debated. Best visited during the calmer spring-to-autumn period when sea conditions are more stable and visibility is more consistent.

Planning Recommendations

• Access: Approximately 30 minutes by air from Naha or Ishigaki airports. The ferry from Ishigaki takes about 4 hours.
• Trip duration: Plan at least 3 nights to account for potential weather cancellations.
• Skill level: Strong currents make this a site for experienced divers. Drift diving proficiency is essential.
• Visibility expectations: Do not rely on historical averages. Conditions can change dramatically with the current. Be prepared for anything from 5 m to 30+ m on any given day.

Scientific Implications and Future Work

Yonaguni's data presents a fundamental challenge to our understanding of ocean visibility prediction. With 4,826 observations, the dataset is more than sufficient in volume. The failure lies not in data quantity but in the available prediction inputs: the physical processes controlling visibility here are simply not captured by standard meteorological and oceanographic reanalysis products. We are exploring several avenues for improvement:

• Integration of high-resolution Kuroshio path data from satellite altimetry
• Addition of Taiwan Strait water temperature and salinity as model features
• Development of upwelling indices based on Yonaguni's submarine topography
• Design of compound predictors capturing tidal-Kuroshio interaction effects

Yonaguni may be the "last frontier" not only for Japanese diving but also for marine data science. If we can crack the prediction problem here, the insights will be transferable to other complex oceanic environments worldwide.