Cross-Regional Marine Ecological Risk Assessment System Integrating Federated Learning

Traditional centralized assessment methods often show bias in global modeling and struggle to balance knowledge sharing with local adaptation when cross-regional data are heterogeneous and non-independent and identically distributed. This problem is common in marine ecological monitoring and is also relevant to distributed electromagnetic sensing networks, where observation standards, sensor density, and local environments vary across regions. To improve the accuracy, consistency, and robustness of cross-regional marine ecological risk assessment, this paper proposes a system incorporating a hierarchical federated learning framework. Each client locally uses a hybrid model integrating physical mechanisms and data-driven capabilities, focusing on phytoplankton-nutrient-dissolved oxygen coupling dynamics. To maintain shared learning consistency, the system applies domain-adaptive regularized aggregation and aligns client hidden-layer distributions on the server, thereby mitigating feature drift. A lightweight personalized calibration layer compensates for regional heterogeneity. Characteristic variables are finally fused into a unified risk scale through uncertainty weighting to produce interpretable results. The method achieves an MSE of 0.483, an MMD of approximately 0.126, and a maximum regional variance of only 0.037, demonstrating stronger cross-regional stability and accuracy than baseline methods while preserving privacy. The system offers quantitative decision support for marine governance and provides a transferable framework for distributed sensor and electromagnetic monitoring applications.