Sea Surface Temperature Prediction Using Machine Learning Techniques

Sea surface temperature (SST) is a fundamental parameter of the Earth system, governing exchanges of heat, moisture and momentum between ocean and atmosphere. Accurate SST prediction is vital for weather forecasting, climate modelling, marine ecosystem management and disaster prevention. Traditional numerical models rely on physical equations and often struggle with computational cost and parameterisation errors, especially at high spatial and temporal resolution. Over the past decade, machine learning techniques have emerged as powerful alternatives or complements, exploiting large‐scale satellite observations and in situ measurements. Approaches range from classical regression and tree‐based methods to deep neural networks that integrate convolutional layers for spatial feature extraction and recurrent architectures for temporal dynamics. Hybrid models combining convolutional neural networks (CNNs) with recurrent units such as long short‐term memory (LSTM) or gated recurrent units (GRU) capture both multiscale spatial correlations and long‐range temporal dependencies. Recent advances also include encoder–decoder frameworks, attention mechanisms and optimisation of network architectures through evolutionary algorithms. These developments have led to marked improvements in short‐term forecasts (days to weeks) and emerging capabilities for seasonal to annual lead times. In parallel, inclusion of auxiliary data—such as wind stress, sea‐level anomalies and regional classifications—has further enhanced prediction skill. As machine learning models become more interpretable and physically informed, they promise to complement classical ocean models and deliver rapid, high‐resolution SST forecasts for diverse applications, from coastal aquaculture to global climate assessment.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Research from all publishers

Recent studies have demonstrated the critical role of spatial information in SST forecasting. A comprehensive evaluation compared linear regression, decision trees, support vector machines and classical neural networks against LSTM and GRU approaches, both with and without explicit spatial inputs. Results on satellite‐derived SST showed that incorporating spatial context reduced prediction errors by approximately 25% and that LSTM models achieved the lowest root mean square error (RMSE) across all metrics. Building on these insights, a dense dilated convolutional LSTM model (D2CL) introduced dilated convolutional layers to extract multiscale spatial features and dense connections to mitigate information loss. Tested on real‐world datasets, D2CL outperformed existing benchmarks for up to seven‐day SST forecasts, demonstrating improved stability and accuracy. Another avenue integrates deep gated recurrent units with convolutional neural networks in a compact architecture optimised via a differential evolution algorithm. Applied to regional SST data in the East China and Yellow Seas, this approach achieved over 98% overall prediction accuracy, with mean absolute errors below 0.33 °C, and outperformed comparative methods in both short‐term and flexible‐horizon forecasting tasks.

Sea Surface Temperature Prediction Using Machine Learning Techniques publication trend

The graph below shows the total number of articles in sea surface temperature prediction using machine learning techniques across all publications each year (not limited to Nature Index journals).

Technical terms

Sea surface temperature (SST): The temperature of the top layer of the ocean, measured by satellites or in situ instruments.

Machine learning (ML): Computational methods that learn patterns from data to make predictions or decisions without explicit physical programming.

Convolutional neural network (CNN): A deep learning architecture that employs convolutional layers to capture spatial hierarchies in input data.

Recurrent neural network (RNN): A class of neural networks designed to process sequential data by retaining information across time steps.

Long short‐term memory (LSTM): An RNN variant that mitigates vanishing gradient issues and retains long‐range temporal dependencies via gated cells.

Gated recurrent unit (GRU): A streamlined RNN cell similar to LSTM that combines input and forget gates for efficient temporal modelling.

Spatiotemporal modelling: The integration of spatial and temporal information to capture evolving patterns in datasets.

Root mean square error (RMSE): A standard metric quantifying average magnitude of prediction errors, penalising larger deviations.

Encoder–decoder framework: A neural network design where an encoder compresses input features into a latent representation and a decoder reconstructs the target output sequence.

References

1. Assessment of the spatiotemporal prediction capabilities of machine learning algorithms on Sea Surface Temperature data: A comprehensive study. Engineering Applications of Artificial Intelligence (2023).
2. D2CL: A Dense Dilated Convolutional LSTM Model for Sea Surface Temperature Prediction. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing (2021).