How Accurately Can AI Predict Diving Visibility?

At IOP, our AI is off by only 2.5m. At Yonaguni, it's useless. Same model, wildly different accuracy. We investigated why AI excels at some sites and fails at others.

What Is the AI Prediction Accuracy Score?

We use a prediction accuracy score (0-100%) to evaluate model accuracy. Values closer to 100% indicate higher prediction accuracy. AI accuracy 82% means the model explains 82% of the variation in actual visibility measurements.

For natural phenomena, an accuracy score above 50% is generally considered useful, and above 70% is considered highly accurate. Given that weather forecast temperature predictions achieve roughly 95%, an accuracy of 82% for something as complex as underwater visibility -- influenced by dozens of interacting factors -- is remarkably good.

Prediction Accuracy by Site

Izu Oceanic Park: A High-Accuracy Model at AI Accuracy 82%

The highest prediction accuracy was achieved at Izu Oceanic Park (IOP) with AI accuracy 82%. This is backed by an extensive dataset of 3,151 observations. IOP logs visibility data nearly every day, providing ample training data for the model. The site also has a clear seasonal visibility pattern -- high in winter, low in spring (a phenomenon known as the spring bloom) -- which the model learns effectively.

The most important predictors for this model are the previous day's visibility, water temperature changes, and the satellite-derived diffuse attenuation coefficient (Kd490). In particular, water temperature shifts associated with the approach of the Kuroshio Current are strongly linked to visibility changes, and capturing this relationship is key to the model's high accuracy.

Ito: AI Accuracy 56%

Ito in Tateyama, Chiba Prefecture achieved AI accuracy 56% with 1,980 observations. Located at the mouth of Tokyo Bay, it sits in a complex marine environment influenced by both the Kuroshio and Oyashio currents. Despite this complexity, its relatively stable visibility patterns allow the model to perform well.

Akinohama: AI Accuracy 52%

Akinohama on Izu Oshima island achieved a moderate AI accuracy of 52% from 1,309 observations. This site maintains relatively stable visibility throughout the year (13-15 m). Because the variation range is small, the model's predictions tend to stay close to the baseline, which naturally results in a more conservative accuracy score.

Kushimoto: AI Accuracy 42%

Kushimoto has one of the largest datasets at 3,168 observations, yet its AI accuracy remains at 42%. While the Kuroshio Current's meandering path strongly influences visibility here, the current's trajectory changes over weeks to months -- a complex phenomenon that same-day weather data alone cannot fully capture. Integrating Kuroshio path data could significantly improve accuracy.

Prediction Intervals via Prediction Range

In addition to single-value predictions, this site also provides a prediction range. This approach calculates a lower bound (10th percentile) and upper bound (90th percentile), providing range-based forecasts such as "visibility will fall between 8 and 15 m with 80% probability." Single-value predictions alone can feel like the prediction "missed," but pairing them with a predicted range makes the information far more practical.

Future Directions

The current models predict same-day conditions and do not yet support multi-day forecasts. Going forward, we aim to achieve 2-3 day visibility forecasting by improving time-series models (GRU, LSTM), integrating Kuroshio path data, and leveraging real-time satellite observations.

We also plan to add more sites to the prediction system as observation data continues to accumulate.