[Predicting soil property in hilly regions by using landscape and multiscale micro-landform features]

Area of Science:

• Soil Science
• Geomorphology
• Remote Sensing
• Machine Learning

Background:

• Accurate digital soil mapping (DSM) is crucial for managing soil resources, especially in complex hilly terrains.
• Traditional DSM methods often struggle with the spatial variability inherent in small watersheds.
• Integrating landscape and micro-landform features offers potential for enhanced prediction accuracy.

Purpose of the Study:

• To assess high-resolution digital soil mapping methods for hilly areas.
• To evaluate the contribution of landscape classification and multiscale micro-landform features in predicting soil properties.
• To compare machine learning techniques for soil property prediction.

Main Methods:

• Utilized Geomorphons (GM) for terrain classification to create landform units.
• Combined traditional DEM derivatives and remote sensing variables with landscape/micro-landform data.
• Employed machine learning algorithms: Support Vector Machine (SVM), Partial Least Squares Regression (PLSR), and Random Forest (RF).
• Applied regression kriging to model residuals of the best-performing prediction models.

Main Results:

• Landscape and micro-landform variables significantly improved prediction accuracy for pH (18.8%), soil clay content (SCC) (8.2%), and cation exchange capacity (CEC) (8.7%).
• Vegetation coverage data within landscape classification showed higher model contribution than land use data.
• A 5 m resolution GM map was optimal for high-precision DSM.
• The Random Forest (RF) model excelled in predicting SCC, but residual regression kriging was not beneficial for pH and CEC predictions with RF.
• A combined approach using landscape, micro-landform, DEM derivatives, and remote sensing variables yielded the highest prediction accuracy for all soil properties.

Conclusions:

• Multivariable data integration provides more comprehensive soil information than single data sources for rolling terrain.
• Landscape variables, including GM and surface classification, explained approximately 40% of the spatial variation in soil attributes.
• Multi-resolution GM and landscape classification variables are valuable components for constructing robust soil mapping prediction models in hilly regions.