A GSD-driven approach to deriving stochastic soil strength parameters under hybrid machine learning models

Abstract

The quantification of soil strength parameters is a crucial prerequisite for constructing physical models related to hydro-geophysical processes. However, due to ignoring soil spatial variability at different scales, traditional parameter assignment strategies, such as assigning values depending on land use classification or other classification systems, as well as those extrapolation and interpolation methods, are insufficient for physical process modelling. This work addressed this deficiency by proposing a method to derive stochastic soil strength parameters under hybrid machine learning (ML) models, taking into account the grain-size distribution (GSD) of soil with scaling invariance. The nonlinear connection between GSD parameters (derived from GSD curves, such as μ and D_c), moisture content, and soil shear strength parameters was fitted by the suggested hybrid ML model. An analysis of a case study revealed that: (i) the Multi-layer Perceptron optimized by the African Vulture Optimization Algorithm (AVOA) algorithm performs the best and can estimate the shear strength parameters of soil mass on vegetated slopes; (ii) all the selected ML models showed significant improvements in predictive performance after optimization with the AVOA, with R² scores increasing by 24.72% for Support Vector Regressor, 34.04% for eXtreme Gradient Boosting, and 35.53% for Multi-layer Perceptron; and (iii) soil cohesion has an increasing relationship with the GSD parameter μ, while soil internal friction angle has a negative correlation with the grain-size parameter D_c. The proposed methodology can give predictions of soil shear strength distribution parameters, providing parameter support for the physical modelling of surface process dynamics.