A comparative study of empirical and machine learning approaches for soil thickness mapping in the Joshimath region (India)

Abstract

Precisely determining the thickness of soil, which is an essential parameter in environmental modelling, presents difficulties when applied to heterogenic large-scale areas. Current prediction models primarily concentrate on shallow soil depths and lack comprehensive spatial coverage. This study addresses this limitation by presenting the results of soil thickness assessment along three important roads in the Joshimath region (Indian Himalaya). Three different methods were examined incorporating geological and geomorphological data as input to obtain soil thickness maps: (1) a customized version of the conventional geomorphologically indexed soil thickness (GIST) model, modified specifically for the peculiarities of the research area, (2) the GIST model enhanced by Monte Carlo simulations (GIST-MCS), and (3) the random forest (RF) algorithm integrated with the GIST model (GIST-RF). By quantifying their errors and conducting validation using geophysical tests, the effectiveness of the models was assessed. Moreover, a critical comparison of the results provided useful insights to understand the peculiarities of the test site and how to adapt the site-specific customization of the models to the local features. The results indicate that the GIST model inadequately accounted for the substantial spatial variations in soil thickness observed across the study area. This is evident from the root-mean-square error (RMSE) of 5.28 m and the mean absolute error of 3.94 m. In contrast, the GIST-MCS model showed improvements, achieving an RMSE of 4.48 m and a mean absolute error of 2.86 m. However, the GIST-RF model demonstrated superior performance, yielding an RMSE of 2.40 m and a mean absolute error of 1.64 m. From a practical perspective, the generated soil thickness maps are particularly significant because they may serve as a crucial input parameter for further studies, including geotechnical assessments, environmental modelling, and decision-making processes. © 2024 Elsevier B.V.