Comparative Performance Analysis of CNN, Linear Regression, and Polynomial Interpolation Models for Sea Surface Temperature Prediction Across the Indian Ocean, Arabian Sea, and Bay of Bengal

ABSTRACT

Accurate prediction of Sea Surface Temperature (SST) is vital for understanding ocean–atmosphere interactions, forecasting extreme weather events, and assessing climate change impacts, especially in tropical regions like the Indian Ocean, Arabian Sea, and Bay of Bengal. In this study, we evaluate and compare the performance of three SST prediction models: Convolutional Neural Networks (CNN), Polynomial Interpolation (Polyfit), and Linear Regression (LR). Using historical MODIS SST data from 2002 to 2010, the generated synthetic SST forecasts from 2011 to 2023 and assessed model performance over three oceanic regions. The CNN model was trained on 192 time steps, capturing complex non‐linear SST dynamics. The LR and Polyfit models provided interpretable and computationally efficient baselines. Results showed that the Linear Regression (LR) model achieved the best performance in the Bay of Bengal (RMSE = 0.70) and the Arabian Sea (RMSE = 0.87), whereas the Polyfit model achieved the lowest RMSE in the Indian Ocean (RMSE = 0.90). Overall, the CNN model demonstrated consistent performance across all three ocean basins, while Polyfit showed relatively poor performance in the Bay of Bengal and Arabian Sea because of its higher prediction errors. All models revealed an increasing SST trend consistent with regional ocean warming. Our findings suggest that while LR is effective in capturing seasonal SST trends, CNNs offer better adaptability to dynamic and non‐linear ocean systems. Polyfit, in contrast, is limited in its ability to generalize over long‐term forecasting. This study highlights the need for region‐specific model selection and advocates for hybrid frameworks that integrate physical drivers and data‐driven learning to enhance SST prediction skill in climate‐sensitive regions. It is important to note that the extended dataset (2011–2023) is synthetically generated and does not represent independent observational validation; therefore, model performance is interpreted as a comparative assessment under controlled conditions rather than a fully empirical forecast validation. Unlike previous studies focusing on single‐model applications, this study introduces a comparative, multi‐basin evaluation framework that highlights region‐specific model suitability under controlled data conditions. The novelty lies in demonstrating that model performance is strongly dependent on oceanographic context, providing a practical basis for selecting appropriate forecasting approaches rather than proposing new algorithms.