AD ‐ GAT : A Decoupled Graph Attention Network for Anisotropic 3D Soil Modeling

Three‐dimensional precision mapping of black soil is a prerequisite for scientifically assessing its degradation status. However, formation and depositional processes determine its strong spatial anisotropy characteristics. Coupled with the sparseness of field borehole data, existing three‐dimensional modeling methods are unable to accurately capture its spatial structure. Given that typical black soils are physically characterized by a surface layer enriched in fine particles (loam), using three‐dimensional texture classification to indirectly represent their spatial occurrence provides a practical and meaningful approach. In response to this challenge, this paper proposes a new deep learning model called the Anisotropic Decoupled Graph Attention Network (AD‐GAT), to model the three‐dimensional stratified structure of soil texture under sparse data conditions. This method uses a parallel architecture to extract content features and incorporates a decoupled spatial attention module (DSAM) to distinguish horizontal and vertical spatial dependencies, thereby learning complex anisotropic patterns. Taking the 3D modeling of Haicheng City as an example (115 boreholes), the superiority of the AD‐GAT model was verified through a systematic comparison with geostatistical methods (3D‐IK), classical machine learning (RF), and standard graph neural networks (GCN). The results show that AD‐GAT achieves superior performance in terms of the Macro F1 score (85.32%) and recall (91.58%), while also attaining a higher Kappa coefficient (0.80). The Monte Carlo Dropout‐based indicators also showed relatively low predictive uncertainty. Ablation experiments further demonstrate the synergistic contributions of individual model components to overall performance improvement. The 3D visualization results further confirm that AD‐GAT can consistently reproduce soil structures with strong spatial continuity and clear stratigraphic boundaries. On this basis, the proposed method provides an effective solution for three‐dimensional soil modeling under sparse data conditions and, by characterizing the three‐dimensional spatial structure of soil texture, offers structural insights for understanding the spatial distribution of black soils and their associated degradation processes.