How to Predict Seawater Temperature for Sustainable Marine Aquaculture (Student Abstract)

Masahito Okuno, Takanobu OtsukaAAAI 2020 · DOI

Introduction

As global demand for seafood continues to rise, marine aquaculture has emerged as a critical industry for food security. However, the industry is highly vulnerable to environmental fluctuations, particularly "abnormal" water temperature shifts that can devastate sensitive crops such as pearls. Traditionally, managing these risks relied on the "gut feeling" and decades of experience of veteran workers. In Japan, as the workforce ages and this institutional knowledge disappears, there is an urgent need for automated, high-precision temperature prediction systems.

Existing forecasting methods often focus on sea surface temperatures or very short-term (next-day) outlooks. In practice, aquaculture products are typically cultivated at depths between 2 and 10 meters. To be useful for farmers, a system must predict temperatures at these specific depths with an error margin of less than 1 °C, allowing for timely preventive action.

Methodology

This study proposes a seawater temperature prediction model utilizing Gated Recurrent Units (GRUs), a specialized type of Recurrent Neural Network (RNN) designed for time-series data. The model architecture is designed to capture the complex relationships between surface conditions and deeper water layers.

Architecture of the GRU-based seawater temperature prediction model.

The model processes two primary data streams:

  1. Seawater Temperature Data: Measurements from multiple depths (0.5m, 2.0m, 5.0m, and 8.0m) are used to account for thermal inertia and vertical heat transfer.
  2. Meteorological Data: Air temperature, wind speed, and temporal markers (month and hour) are included to capture the external forcing factors that influence the water column.

The system is divided into two specialized tracks: a short-term model for hourly predictions over the next 24 hours (to detect sudden spikes) and a long-term model for daily predictions over the upcoming week (for general management).

Study Area and Data

The research utilized a robust dataset spanning a decade (2007–2017) from four major pearl aquaculture sites in Mie Prefecture, Japan: Gokasho Bay, Matoya Bay, Ago Bay (back), and Ago Bay (middle). Water temperature was measured continuously at various depths and transmitted via cellular networks. Nine years of this data were used for model training, with the final year reserved for accuracy verification.

Results and Discussion

The GRU-based approach demonstrated superior performance compared to traditional baseline methods such as Random Forest (RF) and Multiple Regression (MR).

Comparison of Mean Absolute Error (MAE) between GRU and Random Forest (RF) models across different depths and locations.

Key performance metrics include:

  • Short-term Accuracy: The GRU achieved a Mean Absolute Error (MAE) of 0.2 to 0.4 °C, significantly outperforming the Random Forest method, which hovered around an error of 1.0 °C.
  • Long-term Accuracy: For one-week-ahead forecasts, the GRU maintained an MAE of 0.2 to 0.5 °C, consistently beating Multiple Regression results by approximately 0.1 °C.

A critical insight from the study is the role of thermal inertia; because seawater temperature changes more slowly than air temperature, feeding multiple depth layers into the GRU allows the model to learn the specific "lag" and vertical relationships within the water column. Since the error rates are roughly half of the 1 °C threshold required by pearl farmers, the system is considered ready for practical deployment.

Conclusion

The proposed GRU model provides a reliable, data-driven alternative to traditional experience-based management in aquaculture. By maintaining high accuracy at specific cultivation depths, the system empowers new workers to manage environments as effectively as veterans. Future work will focus on generalizing the model to other geographical regions and further refining the early warning capabilities to protect marine stock from climate-driven temperature anomalies.

Summary updated

BibTeX
@inproceedings{okuno2020how,
  author = {Masahito Okuno and Takanobu Otsuka},
  title = {{How to Predict Seawater Temperature for Sustainable Marine Aquaculture (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7216},
}
Figure