Linking hydroclimatic hazards and catchment vulnerability to river ecological status under a data-sparse condition using hybrid graph neural networks

River water quality status is increasingly challenged by the combined effects of climate extremes and human activities. Furthermore, monitoring water quality parameters remains sparse and irregular across many river networks to conduct timely and effective assessments. To address these challenges, a new framework was developed to support river network-wide estimation of ecological water quality status. The framework combines a multi-risk approach with deep learning to investigate the relationship between climate hazards, anthropogenic pressures, exposure, and vulnerabilities under hot-dry and wet-dry conditions. It was implemented for the Veneto Region (Italy) to predict annual LIMeco, a nutrient-oxygen physico-chemical index used in regional assessments, for 865 river segments over 2010-2023, with data missingness reaching 66.8%. Hazard conditions were represented by annual hydroclimatic indicators capturing hot and wet/dry conditions and extremes, while anthropogenic pressures were described through land use composition and nutrient load proxies. Exposure and vulnerability were represented through basin characteristics (e.g., soil properties and topography), together with the presence of riparian and wetland areas as proxies of buffering capacity and management levels. To translate these drivers into spatially coherent predictions while acknowledging missing observations due to incomplete datasets, a hybrid spatio-temporal Graph Neural Network (GNN) was implemented in which (i) recent hydroclimatic variability was summarized over a short input window using the Gated Recurrent Units, (ii) information was propagated along upstream-downstream connectivity using GNN, and (iii) eco-hydrological clusters of basins were represented through a data-driven regime label with an unsupervised Machine Learning (ML), derived from the multi-risk indicators, enabling the information transfer between well-monitored and poorly-monitored river segments that have similar climate and land-based regimes. The hybrid spatio-temporal GNN was tested over multiple configurations and against other ML approaches (i.e., Multilayer perceptron and eXtreme Gradient Boosting). The comparison demonstrates that the hybrid GNN achieved the best performances with the highest test accuracies  (RMSE = 0.028; NSE = 0.98) when using the embedded river basin clusters, providing the most stable performance across basin types and missingness levels. This highlighted how the inclusion of physically-based river network dynamics and auxiliary information can help to address missing data compared to other tested methodologies. The proposed framework can support scenario-oriented analysis for decision making and planning, given the representation of management-relevant indicators (e.g., riparian condition, wetland presence, land-use pressure), allowing for exploring future scenarios and responses under climate and socio-economic changes to support adaptation strategies.