Marine Heat Waves (MHWs) are extreme ocean temperature anomalies that can disrupt marine ecosystems, fisheries and coastal economies. Early and accurate prediction of MHWs is critical to support environmental monitoring and effective mitigation strategies. In this paper, we propose a novel federated learning framework for distributed prediction of MHWs using Sea Surface Temperature (SST) data collected from in situ sensors located along the Italian coastline. Our approach leverages the decentralized nature of marine monitoring infrastructures, allowing each coastal station to train local models in site-specific SST time series without sharing raw data, thus preserving data privacy and compliance with data sovereignty regulations. The system employs two LSTM architectures, used with FedAvg and personalized federated learning strategies to collaboratively aggregate local models. The collaborative federated learning paradigm improves the predictions of SST and MHW by effectively capturing distributed regional dynamics. Experimental results show that the proposed federated learning approach outperforms local on-site training (average RMSE over 1 to 7 day forecasts: 0.89  \(^{\circ }\) C vs 1.11  \(^{\circ }\) C) and almost matches the accuracy of centralized training, which assumes access to all raw data from every site (0.82  \(^{\circ }\) C). Our work lays the foundation for a scalable and privacy-aware digital infrastructure for climate resilience in marine environments.