Advancing Riverine–Lacustrine Ecosystem Vulnerability Prediction Using Multi-Sensor Satellite Data, Attention-Based Deep Learning, and Evolutionary Metaheuristics

Riverine–lacustrine ecosystems in river–lake continua face increasing threats, yet conventional vulnerability maps often overlook local degradation drivers. This study presents an advanced satellite-based mapping framework using Deep Attention Networks (DANets) for accurate, interpretable vulnerability assessment. In the Ebinur Lake Basin, a representative dryland river system, we first built a satellite-derived evidence map of ecosystem stress aligned with the IPCC’s vulnerability definition. We then optimized DANets via two nature-inspired algorithms: Genetic Algorithm (GA) and Grey Wolf Optimizer (GWO). The optimized models demonstrated strong predictive capacity, explaining a large share of vulnerability variance (R2 = 0.78 for GA-DANets; R2 = 0.76 for GWO-DANets). For high/low-vulnerability discrimination, GWO-DANets was most effective and stable, with a mean AUC = 0.960 ± 0.044. Factor importance analysis identified soil organic carbon (SOC; 0.29), precipitation seasonality (0.24), and aridity (0.22) as dominant drivers. Two distinct pathways emerged: chronic degradation in arid plains, driven by low SOC and poor water retention; and acute hydrological stress in wetlands, where carbon-rich soils are sensitive to drying. This insight shifts management from uniform to targeted approaches: soil restoration in plains and water-flow protection in wetlands. By integrating metaheuristically optimized deep learning with multi-sensor satellite data, the framework offers a scalable decision-support tool for safeguarding water-dependent ecosystems. The study confirms that vulnerability in the basin follows two predictable, process-based trajectories, which can be directly linked to measurable soil and hydrological conditions. These clear patterns allow managers to prioritize interventions where they will have the greatest effect under ongoing climate pressure.