Published Paper Details:

This study develops a hybrid machine learning framework to analyze and forecast surface ocean dynamics in the Arabian Sea. The pipeline integrates preprocessing, anomaly detection, and forecasting within a cycle-wise modular design that separates carbon, nutrient, phytoplankton, oxygen, and physical drivers into independent modeling workflows. Classical ensemble methods such as support vector machines and clustering were employed for anomaly detection, while hybrid CNN-LSTM architectures were trained for temporal forecasting, supported by additional ensemble models (ARIMA, Prophet). Evaluation demonstrated that the modular design is computationally feasible and ecologically interpretable, successfully reproducing key surface-layer dynamics including carbonate variability and nutrient-chlorophyll coupling. Carbon cycle forecasts achieved the strongest performance, while oxygen predictions were more uncertain due to gradient complexity and data sparsity. Although constrained by reduced training and absence of in-situ validation, the framework establishes a baseline methodology for hybrid ML-DL applications in marine forecasting, offering a scalable approach for future integration with physical ocean models.