Comparative ocean temperature forecasting using SARIMA, ETS and LSTM

Abstract

Accurate ocean temperature forecasting is essential for understanding long-term environmental variability and supporting ecological decision-making. This study evaluates the performance of SARIMA (Seasonal Autoregressive Integrated Moving Average), ETS (Error Trend Seasonality), and LSTM (Long Short-Term Memory) models for predicting ocean temperature using historical time series data from the CalCOFI programme. The dataset was preprocessed using seasonal decomposition and stationarity analysis, and forecasting accuracy was assessed using MAPE (Mean Absolute Percentage Error), RMSE (Root Mean Squared Error), and MSE (Mean Squared Error). The results indicate that LSTM produced the most accurate and stable forecasts overall, achieving lower RMSE and MAPE values at most stations and depths. It effectively captured nonlinear behavior, seasonal variability, and extreme temperature fluctuations. SARIMA demonstrated a strong capability to model trend and seasonality, while ETS generated smoother forecasts but with comparatively higher error values. Residual diagnostics further confirmed LSTM’s superior ability to learn complex temporal dependencies present in ocean temperature data. This study contributes to a comparative evaluation of classical statistical and deep learning models for ecological time series forecasting and provides evidence supporting the application of LSTM for improved oceanographic prediction and environmental monitoring.