Assessing soil thickness and distribution in subtropical typhoon areas: an integration of advanced geomorphological surveys and ensemble learning approaches

Background: In subtropical-typhoon regions, prolonged rainfall often triggers landslides, creating challenges in predicting soil thickness and its spatial distribution due to dense vegetation and complex topography. This study aims to tackle these issues by developing a watershed-scale map of unstable layer thickness to improve predictions of debris flows and landslides. Methods: We integrated geomorphological surveys with ensemble machine learning techniques. Unmanned Aerial Vehicle technology was used to create a 3D digital elevation model, identifying different Quaternary deposits in the study area. Fieldwork involved three stages: soil thickness data collection via field surveys, geophysical exploration for spatial distribution, and core drilling for geotechnical properties. Separate evaluation systems were built for eluvium and slope deposits. A machine learning model was developed on a Python platform to predict soil thickness, using data from mountainous watersheds along China’s eastern coast. Results: The model accurately predicted 84.7% of eluvium soil thickness (average 0.64 m, mainly sandy clay) and 81.3% of slope deposit thickness (average 2.34 m, including sandy clay and crushed stone). For eluvium, the root mean square error was 0.148 m, and for slope deposits, it was 0.27 m. Key influencing factors were lithology for eluvium and elevation for slope deposits. Shallow landslides were most prevalent in these layers, with sliding surfaces at specific interfaces between material types. Conclusions: This study demonstrates the effectiveness of combining geomorphological surveys and machine learning for precise soil thickness prediction. The methodology enhances geohazard models, offering insights into landslide behavior and supporting more accurate risk assessments. These findings provide a foundation for future research on mitigation strategies in similar regions.