Quantifying the effects of data preprocessing and location-level aggregation on SPT-based machine learning prediction of safe bearing capacity: evidence from the Godavari-Krishna Delta, Andhra Pradesh, India

Accurate estimation of Safe Bearing Capacity (SBC) is essential for shallow foundation design, yet empirical correlations with Standard Penetration Test (SPT) N-values are largely univariate and perform poorly in geologically variable deltaic regions. Two challenges limit machine learning (ML) alternatives: the benefit of systematic preprocessing remains unmeasured, and the depth-wise structure of borehole datasets introduces autocorrelation that can bias evaluation. This study presents a Random Forest (RF) and Gradient Boosting (GB) framework isolating these challenges via a controlled 2×2 design, applied to 4,526 SPT records from 838 boreholes in the Godavari-Krishna delta, India. Preprocessing reduced the dataset to 3,950 records and decreased RF and GB RMSE by 17.4% and 17.1%, while aggregation reduced RMSE by 32.5–38.6%. The best model (GB, processed) achieved R2 = 0.941, RMSE = 1.905 t/m2; the RF Aggregated model (R2 = 0.938, RMSE = 1.290 t/m2) is recommended for practical design. SHAP confirmed N-value as the dominant predictor (71–85%), with the top three features jointly accounting for 88–96%, and empirical methods yielded negative R2, with ML reducing RMSE by 90–95%. A linear regression baseline achieved R2 = 0.902, only marginally below the best model, confirming that data quality and structure, not algorithm choice, drive predictive reliability.