Dual-Channel Temporal Graph Attention Network for Forecasting Sodium Hazard in Irrigation Water Systems

Groundwater quality for irrigation is increasingly threatened by overextraction, salinization and agricultural intensification, yet existing sodium hazard prediction methods often fail to capture complex temporal and relational dynamics. This study proposes a Dual-Channel Temporal Graph Attention Network (DC-TGAN), integrated with advanced preprocessing and hybrid feature selection, to enhance the prediction of farming water quality characteristics—Kelly's ratio (KR), residual sodium carbonate (RSC), sodium adsorption ratio (SAR), and %sodium (%Na)—in Southern Rajasthan, India's Pratapgarh district. The approach's originality is found in its ability to simultaneously model multivariate spatiotemporal dependencies and optimize convergence through Quokka Swarm Optimization. The model was found to best predict groundwater when using 10 years (2010–2020) of groundwater data, with correlation coefficients (R2) near to 1.0 and testing RMSE of 0.04 in KR and 0.16 in SAR. On the whole, the DC-TGAN was superior to the models of random forest, ANN, LSTM and WDO-ANN with an accuracy of 99.8%, F1-score of 99.7%, and the T of 0.02 s. These findings indicate that the model can be viewed as a powerful and scalable instrument of timely identifying sodium hazards to contribute to sustainable groundwater management and irrigation planning in semiarid areas.