Predicting active-layer soil thickness using topographic variables at a small watershed scale.

Area of Science:

• Geosciences
• Environmental Science
• Soil Science

Background:

• Accurate spatial distribution of active-layer (AL) soil thickness is vital for ecological modeling, precision agriculture, and land management.
• Traditional soil survey methods struggle to provide detailed AL soil thickness data.
• Terrain variables offer a potential alternative for mapping AL soil thickness.

Purpose of the Study:

• To investigate the feasibility and accuracy of mapping AL soil thickness spatial distribution.
• To evaluate the performance of a random forest (RF) model using terrain variables at a watershed scale.
• To compare RF model performance against multiple linear regression (MLR), geographically weighted regression (GWR), and support vector machines (SVM).

Main Methods:

• Collected 1113 soil samples, divided into calibration (770) and validation (343) sets.
• Derived seven terrain variables (elevation, aspect, etc.) from a 30m digital elevation map (DEM).
• Employed RF, MLR, GWR, and SVM models, evaluating performance using ME, MAE, RMSE, and R2.

Main Results:

• The RF model demonstrated superior performance compared to MLR, GWR, and SVM.
• RF achieved ME of 0.39 cm, MAE of 7.09 cm, RMSE of 10.85 cm, and R2 of 62%.
• Elevation, flow path length, and valley depth were identified as key factors influencing AL soil thickness.

Conclusions:

• The RF model is a promising method for predicting AL soil thickness spatial distribution using terrain parameters.
• Terrain variables derived from DEM are valuable inputs for AL soil thickness mapping.
• The study highlights the potential of machine learning for soil science applications.