HydroFinNet: An Explainable Multi-Modal Deep Learning Framework for Sustainable Water Resource Allocation and Economic Risk Assessment

Sustainable water resource management requires intelligent decision-support systems capable of addressing hydrological uncertainty, competing sectoral demands, and associated economic risks. Existing approaches commonly rely on isolated forecasting or optimization models and provide limited interpretability for critical allocation decisions. This paper proposes HydroFinNet, an explainable multi-modal deep learning framework for jointly optimizing sustainable water allocation and assessing economic risk. The proposed architecture integrates remote-sensing and reservoir observations, meteorological and hydrological time series, demographic and sectoral demand indicators, and financial market variables through modality-specific encoders and a cross-modal multi-head attention mechanism. A dual-task architecture simultaneously generates sector-wise allocation recommendations and an Economic Risk Index while incorporating ecological-flow constraints into the optimization objective. To improve transparency, Deep-SHAP attribution and attention-based analysis identify influential features and modalities underlying model decisions. Experimental evaluation demonstrates improved water-demand prediction, allocation efficiency, and economic-risk estimation compared with conventional statistical and uni-modal deep learning baselines. The framework provides an interpretable and scalable foundation for data-driven, risk-aware, and sustainable water-resource governance.