Machine Learning for Subgrade Moisture in Cold Regions: Random Forest and Gaussian Mixture Models

Understanding subgrade moisture content is critical for ensuring the structural integrity and performance of roadways, particularly in cold regions where snowmelt-induced thawing can significantly increase moisture levels. This study aims to develop accurate and practical tools for predicting subgrade moisture variation using both traditional statistical methods and advanced machine learning approaches. The primary goals are to improve prediction accuracy, capture seasonal and temporal variations, and provide highway agencies with reliable tools for decision-making. To achieve these goals, two sites in central New York, USA were instrumented with moisture sensors placed at depths ranging from 0.3 to 1.2 m, and data were collected over three years (May 2019–April 2022). The methodology involved applying a Generalized Linear Model (GLM) as the traditional regression approach and comparing its performance with two machine learning techniques: Random Forest (RF) and Gaussian Mixture Model (GMM). The GLM yielded an R² value of 0.72. In contrast, RF and GMM significantly improved predictive accuracy, achieving R² values of 0.80 and 0.94, respectively, along with notably low mean squared error values. The four-component GMM demonstrated superior performance, effectively capturing the seasonal and temporal moisture variations at different depths. The contributions of this study are twofold: first, it highlights the limitations of traditional regression methods in modeling complex subgrade moisture dynamics; second, it establishes the GMM as a robust and transferable tool for subgrade moisture prediction. The findings are of particular importance for highway agencies, as the proposed approach offers a flexible, data-driven framework to enhance the design, maintenance, and resilience of low-volume roads in cold climates.