With the value enhancement of the data elements in the water resource sector, the requirements for security and compliance in data usage have been increasingly strengthened, bringing challenges to cross-departmental data collaboration. This paper proposes a data governance framework based on federated learning, which improves the efficiency of data utilization among multiple entities through a distributed modeling mechanism while effectively avoiding privacy and compliance risks caused by centralized processing. Prediction of measured runoff data for a representative in the Nanjing reach of tributary the lower Yangtze River, and the results show that, under the premise of ensuring data privacy, the model achieves prediction accuracy significantly superior to that of local independent models and exhibits good generalization performance. The results show that federated learning can effectively break down data barriers while keeping the original data within its domains, transforming dispersed data resources into high-value predictive capabilities. The performance gains remain consistent across different architectures, indicating that its effectiveness stems from the collaborative paradigm. Therefore, federated learning provides feasible technical support for the standardized sharing and intelligent analysis of water resource data, demonstrating good engineering adaptability and promotion potential.