Graph-based and machine learning approaches for soil depth prediction in a reservoir landscape: A case study in Dazhou County, Chongqing, China

Abstract

Reservoir-bank areas are characterized by intense soil erosion and deposition processes, resulting in significant spatial variations in soil thickness that influence landslide occurrence and threaten resident safety. This study presents an adaptive modelling framework to predict soil thickness by capturing the complex spatial relationships inherent in its distribution, significantly improving prediction accuracy. A reservoir-bank area of 1.7 Km² in Dazhou town, Chongqing Province, China, was selected as a study area. A total of 288 soil thickness samples derived from field observation and drilling works, along with 14 environmental factors (such as altitude, slope, relative slope position index (RSPI), and sediment transportation index) were utilized to generate the initial modelling dataset. Subsequently, two graph models were developed based on the feature and geographic similarity, and the extracted graph features were integrated with environmental factors as inputs for machine learning models, including Random Forest (RF), Support Vector Machine, and Gradient Boosting Decision Tree (GBDT), to predict soil thickness maps. The validation results of root-mean-square-error (RMSE), coefficient of determination (R²), and error frequency analysis highlighted two essential conclusions in this study: i) Among the three models, the GBDT model showed the best performance overall, with the highest R² (0.7431 for testing, 0.9569 for training), the lowest RMSE (5.3189 for testing, 2.3001 for training), and the lowest residual skewness value of 0.11. ii) Incorporating graph-based features significantly enhances the accuracy of soil thickness predictions, particularly for nonlinear models (RF and GBDT), by effectively mitigating overestimation issues caused by spatial dependencies among independent variables (such as altitude and RSPI). This study integrates machine learning techniques with graph-based spatial analysis, providing a new path for advancing soil thickness prediction research.