Why AI Can Predict Visibility at Some Sites but Not Others: Pattern Analysis

R²=0.82 vs R²=0.00: Heaven and Hell of Prediction Accuracy

Our AI visibility prediction model achieves an average R²=0.82 across all sites, but accuracy varies dramatically by location. IOP (Izu Oceanic Park) achieves R²=0.824 with high accuracy, while Yonaguni is R²=0.046 — nearly random. Osezaki Bay achieves R²=0.000 — completely unpredictable. Why such dramatic differences? We analyze the relationship between site characteristics and AI prediction accuracy.

Site-by-Site Prediction Accuracy

R² = coefficient of determination. 1.0 = perfect prediction; 0.0 = equivalent to random. — = not measured.

Why Accuracy Differs: Site-Type Analysis

Open Ocean (Yonaguni, Kerama): 'High Clarity but Unpredictable'

Yonaguni's R²=0.046 is shocking. Even combining all weather, marine, and satellite data used by the AI, almost none of tomorrow's visibility can be explained. The cause is the Kuroshio main stream's direct influence. The Kuroshio shifts position subtly day by day — on some days, clear oceanic water delivers 50m visibility; after typhoon turbulence, it can drop to 10m. This 'Kuroshio lottery' cannot be predicted from current weather data alone.

Enclosed Bay (Osezaki Bay): 'Low Clarity AND Completely Unpredictable'

Osezaki Bay's R²=0.000 is literally 'zero prediction accuracy.' Average visibility is low at 7.6m, and what determines it is completely unknown. In enclosed bays, fine-scale factors like internal algal dynamics, sediment movement, and local wind stirring control visibility — factors that broad weather data cannot capture. Even though the AI learns features that work across Japan, it cannot account for bay-specific local dynamics.

Coastal (IOP, Futone): 'Where AI Works Best'

IOP's R²=0.824 is Japan's best. At coastal sites, seasonal plankton cycles, wave-induced turbidity, and previous days'/weeks' visibility (lag features) consistently explain visibility. Strong autocorrelation — 'yesterday was 15m so today will be similar' — allows the AI's lag features to work effectively. Also, buffering from direct Kuroshio and typhoon impacts means fewer sudden changes.

What Determines Predictability

What This Means for Divers

At hard-to-predict sites (Yonaguni, Kerama), AI confidence intervals are wide — predictions often come as broad ranges like '15–25m' rather than precise numbers. At IOP or Akinohama, intervals are narrow (e.g., '13–15m'), making trip planning more reliable.

At hard-to-predict sites, checking real-time conditions right before diving is critical. At sites like Yonaguni where the Kuroshio's daily position determines visibility, the previous day's actual dive log reports from local dive services are the most reliable source of information.

Summary

Visibility predictability is largely determined by a site's geographic characteristics. Open ocean sites (Yonaguni) are hard to predict due to direct Kuroshio influence. Enclosed bays (Osezaki Bay) are completely unpredictable due to local factors. Coastal sites (IOP, Futone) have strong seasonal patterns and autocorrelation where AI performs best. Understanding predictability helps you know how much to trust AI forecasts and make smarter dive planning decisions.

R² values from LightGBM model holdout validation set. Values for sites with few observations are indicative.