Arid socio-ecological systems (SESs), such as the Sistan Watershed, face severe crises including wetland desiccation and recurrent dust storms. Modeling the complex, non-linear dynamics of these systems is a significant challenge due to data scarcity and the opacity of conventional AI models. This study introduces a novel process-informed Graph Neural Network framework to forecast these dynamics. We first reconstructed a 21-year aerosol index time series using a machine learning-based data fusion to address significant data gaps (MNAR). A GNN embedded with the system's causal structure was then trained to predict the future state of key environmental variables. The model achieved robust performance on unseen data (R² = 0.55 for water presence, R² = 0.43 for aerosol index). Scenario simulations demonstrated that wetland restoration can significantly improve the aerosol index (over 55% reduction), offering a powerful tool for sustainable management in complex environmental systems.
Socio-ecological systems (SESs) in arid regions are increasingly vulnerable to climate change and anthropogenic pressures, often approaching critical "tipping points" that threaten ecosystem services and human well-being (Scheffer et al., 2001). The Sistan Basin, a transboundary watershed on the Iran-Afghanistan border, epitomizes this crisis. The desiccation of its Hamoun Wetlands has triggered severe dust storms, impacting public health and regional stability (Karimi et al., 2024). While advanced predictive models are urgently needed, traditional approaches face limitations. Physically-based models struggle with SES complexity, while data-driven AI models often act as "black boxes," learning spurious correlations and lacking interpretability, which limits their utility for policy-making (Reichstein et al., 2019). This research addresses this methodological gap by developing a process-informed, "gray-box" deep learning framework. The primary objective is to create a robust and interpretable model capable of forecasting the dynamics of the Sistan SES and evaluating the impact of different management interventions, thereby providing a quantitative decision-support tool.
Methodology
This study employed a multi-stage framework. First, we conducted data acquisition and fusion. Time-series data (2001-2021) for six key environmental variables (e.g., precipitation, water presence, aerosol index) were extracted from multi-source satellite data using Google Earth Engine. To address a significant gap in the aerosol time series, we implemented a machine learning-based data fusion technique, training a Random Forest Regressor to reconstruct the missing data. Second, we developed a novel process-informed Graph Neural Network (GNN). This model embeds a predefined causal graph, representing known process knowledge, into its architecture. The model combines Gated Recurrent Unit (GRU) layers to capture temporal patterns and a Graph Convolutional Network (GCN) layer to learn systemic interdependencies. The model was trained using the Adam optimizer and an MSE loss function, with an early stopping mechanism to prevent overfitting. Finally, a series of post-hoc analyses were performed, including model performance evaluation, sensitivity analysis, and the simulation of three future scenarios: Baseline, Sustained Drought, and Wetland Restoration.
Results and Discussion
The trained model demonstrated robust predictive performance on the unseen test set (2019-2021), achieving an R² of 0.55 for water presence and 0.43 for the aerosol index. This validates the model's ability to capture the core dynamics of the system. A sensitivity analysis revealed a notable inertia in the Sistan SES, showing a less than proportional response to gradual changes in precipitation (-50% to +50%). This suggests that other drivers, particularly upstream water management, may have a more dominant influence on the system's state.
Scenario simulations yielded critical insights for policy-making. While the drought scenario showed a moderate decline in ecosystem health, the wetland restoration scenario projected a powerful positive cascade. By maintaining water presence at its historical 75th percentile, the model forecasted a significant 55.6% improvement in the aerosol index (indicating a substantial reduction in local dust pollution), alongside an 11.8% increase in wetland vegetation. This result is highly significant as it quantitatively demonstrates that proactive, internal management interventions can be a highly effective lever for mitigating a key environmental hazard. Unlike previous studies that may have focused on single aspects, our integrated SES model reveals the systemic, cascading benefits of such an intervention, moving beyond simple prediction to provide a tool for evaluating policy effectiveness. The findings strongly suggest that hydrological restoration of the Hamoun Wetlands is the most impactful local strategy for combating the regional dust crisis.
Conclusion
This study successfully developed and validated a novel, process-informed deep learning framework for modeling complex, data-scarce socio-ecological systems. The model demonstrated strong predictive performance and, through scenario simulations, provided critical, actionable insights. Our findings quantitatively confirm that the hydrological restoration of the Hamoun Wetlands is a highly effective strategy for mitigating local dust pollution and improving regional ecosystem health. This framework serves as a powerful and interpretable "digital twin," offering a quantitative decision-support tool to guide sustainable management and enhance resilience in the Sistan Basin and other vulnerable arid regions.