A probabilistic pedotransfer function for estimating the soil water characteristic curve and quantifying its uncertainty

Le résumé fourni par la source

The fast and accurate prediction of the soil water characteristic curve (SWCC) is important for geotechnical engineering, but the traditional predictive models are often deterministic, have limited generalizability, and fail to quantify the predictive uncertainty. To address this issue, we propose a novel two-stage framework that integrates the inferential strength of Bayesian Markov chain Monte Carlo (MCMC) with the predictive strength of CatBoost: first, Bayesian MCMC was used to infer the full posterior distributions of the van Genuchten parameters for 1249 soil samples obtained from the UNSODA and HYBRAS databases; second, we used CatBoost regressor to train a machine learning model to predict the statistical characteristics of these posterior distributions from basic soil properties. The best fit curve (i.e., the mean value of parameters from Bayesian inference) of the resulting model demonstrates higher predictive accuracy than the well-recognized HYPRES and ROSETTA models. More importantly, we built a Bayesian uncertainty-quantified pedotransfer function, called BUQ-PTF, which reduced the root mean square error (RMSE) by 40–50% relative to the HYPRES and ROSETTA models and achieved appreciable coverage probability in prediction intervals. This study marks a significant advancement toward the probabilistic prediction of SWCC from a deterministic single-curve prediction, thus providing a robust tool for risk-informed decision making in engineering applications.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.