Knowledge discovery for interpretable and variability-aware permeability prediction under sparse sampling in tropical weathered soils: An Amazon case study

This is the authors' abstract. We don't add key points for this paper.

Permeability assessment in tropical weathered soils is limited by sparse sampling, difficult access, and the high cost of field hydraulic testing. This limitation is critical along transportation corridors in the Brazilian Amazon, where lateritic materials show strong spatial variability and influence slope response. This study proposes a spatial-to-parametric transfer framework for permeability inference in the Itapecuru Formation along the Carajás Railway, Brazil. The framework uses conditional random fields (CRF) to transfer site-conditioned spatial variability into a machine-learning stage, where permeability is related to routine geotechnical descriptors and interpreted through explainability tools. LSSVR, MLP, and XGBoost were evaluated under baseline and optimized configurations. The CRF-derived domain preserved the main statistical structure of the measured data, with non-significant distributional differences for most variables. Hyperparameter optimization improved model behavior, especially for LSSVR and MLP. PSO-LSSVR provided the best measured-reference performance, with RMSE = 1.075 and MAE = 0.811, and showed the most stable response across repeated spatial-fold partitions. The CRF-derived variability intervals supported the diagnosis of prediction difficulty and helped identify observations associated with broader response ranges. SHAP and partial-dependence analyses (XAI) indicated that the inferred permeability response was mainly organized by plasticity-related restriction and granulometric–compositional controls associated with aggregation-sensitive descriptors. These results show that sparse permeability measurements can be extended into an interpretable and variability-aware inference framework, supporting preliminary hydraulic assessment in tropical weathered soils where direct testing is limited.