Enhancing class-imbalanced soil classification using an explainable tabular network and synthetic minority over-sampling framework

Accurate characterization of soil engineering properties is fundamental to effective infrastructure design; however, traditional soil classification approaches are resource-intensive, requiring extensive laboratory testing which affects road project budgets and timelines. Furthermore, geotechnical datasets suffer from severe class imbalances, which bias predictive models toward dominant soil types. This study proposes an explainable deep learning framework centered on Tabular Network (TabNet) via Bayesian Optimization to address soil classification in tabular data. To mitigate data scarcity and class imbalance, Synthetic Minority Over-sampling (SMOTE) was integrated to generate synthetic instances of minority classes. The empirical results demonstrate that for the studied regional soil profiles, the TabNet-SMOTE framework achieves superior predictive performance, with an overall accuracy of 0.95 and Macro F1-score of 0.94, outperforming advanced generative models. A multifaceted performance analysis indicated the model's high ordinal fidelity, with a Quadratic Weighted Kappa (QWK) of 0.95 and negligible Mean Error of 0.036. Furthermore, SHapley Additive exPlanations (SHAP) provides global and local interpretability, revealing that the 0.075 mm sieve fraction and Atterberg limits constitute a critical predictive core, accounting for 83.8% of cumulative feature importance. By aligning data-driven insights with established geotechnical principles, this research provides a statistically robust, transparent, and efficient decision-support framework. In practical applications, the framework enables engineers to input laboratory index properties and obtain soil classifications immediately following index testing, supporting subgrade assessment and pavement design optimization while ensuring decision transparency through SHAP-based interpretability. This substantially enhances the geotechnical workflow by providing rapid, automated classification based on laboratory index properties in data-constrained regions.