Akademik Veri Yönetim Sistemi
EXPERT SYSTEMS WITH APPLICATIONS, cilt.314, ss.1-14, 2026 (SCI-Expanded, Scopus)
Accurate forecasting of subsurface soil temperature is essential for climate-smart agriculture, environmental monitoring, and data-driven resource management. Modern agricultural information systems increasingly depend on predictive analytics, yet conventional physics-based models require extensive calibration, and purely data-driven methods often struggle to generalize across heterogeneous landscapes. To address these limitations, this study introduces a hybrid predictive framework that integrates Physics-Informed Neural Networks (PINNs) with Graph Neural Networks (GNNs) for use within decision-support and environmental information systems. The PINN component embeds the one-dimensional heat-diffusion equation into the learning process, producing physically coherent temperature profiles without the need for detailed soil parameters. In parallel, the GNN models spatial relationships among meteorological stations by encoding geographic and topographic similarity, enabling the system to transfer knowledge across regions and improve generalization.
The model is trained and evaluated using a multi-year dataset from 15 meteorological stations across Türkiye, representing diverse climatic and soil conditions. Comparative experiments show that the hybrid GNN-PINN consistently outperforms standalone PINN, GNN-only, and multilayer perceptron baselines. Improvements are most pronounced at deeper soil layers (50–100 cm), where thermal diffusion dominates and conventional models tend to lose accuracy. Ablation analyses reveal that physics-based constraints and spatial message passing make complementary contributions, with statistically significant performance gains (p < 0.05). Despite these en hancements, the hybrid model requires only ~ 20% additional training cost and maintains near real-time inference.
By combining physical principles with spatially aware learning, the proposed framework offers reliable, interpretable, and computationally efficient soil-temperature forecasts, supporting data-driven decision making in climate-smart agriculture and environmental management.