[A-TT35] Machine Learning Techniques in Weather, Climate, Ocean, Hydrology and Disease Predictions

[ATT35-07] Predicting Characteristics of Salmon Return Migration Using Deep Learning

*Mikhail Borisov¹、Mikhail Krinitskiy^1,2 (1.Moscow Institute of Physics and Technology、2.Shirshov Institute of Oceanology, Russian Academy of Sciences)

キーワード：Artificial Intelligence, Oceanography, Sockeye Salmon, Salmon Return Migration

In the Fraser River mouth region of British Columbia, the commercial fishing of sockeye salmon by Canadian and American entities takes place within a network of straits between Vancouver Island and mainland Canada. The timing of fishing seasons and the allocation of sockeye salmon catches between the USA and Canada are influenced by the characteristics of the salmon's return migration through these straits. This study hypothesizes that sockeye salmon behavior is affected by factors such as seawater chemistry, sea surface temperature, and ocean current dynamics. The research utilizes a methodology incorporating contrastive learning with Momentum Contrast (MoCo) and Long Short-Term Memory (LSTM) networks for the dimensionality reduction of 2D oceanic data. Following this, traditional machine learning models, including linear regression, Ridge regression, and Random Forest, are employed to predict the northern diversion rate (NDR) and the ordinal date of median return (RT) of salmon migration. The training dataset is derived from Glorys12 reanalysis data, which includes variables such as surface current velocities, salinity, and other environmental factors. Previous efforts to predict NDR and RT have either focused on pure autoregression or statistical modeling using environmental predictors. However, this study marks the first attempt to leverage deep learning, specifically self-supervised pretraining, alongside advanced classical machine learning models, to address the prediction of salmon behavior. The study details the data processing and model training procedures applied, which present opportunities for enhanced prediction accuracy and insights into salmon migration dynamics in the complex Pacific Ocean system.