Digital soil mapping including additional point sampling in Posses ecosystem services pilot watershed, southeastern...

Area of Science:

• Environmental Science
• Soil Science
• Geospatial Analysis

Background:

• Digital soil mapping relies on spatial association models to predict soil properties.
• Steep-slope watersheds present unique challenges for accurate soil mapping due to complex terrain and potential data imbalance.
• Evaluating model performance and the impact of sampling strategies is crucial for improving soil survey accuracy.

Purpose of the Study:

• To assess the performance of Multinomial logistic regression (MLR), C5 decision tree (C5-DT), and Random Forest (RF) models in digital soil mapping.
• To investigate the effect of additional point sampling, obtained via photointerpretation, on model accuracy.
• To evaluate the extrapolation accuracy of the models using Multivariate Environmental Similarity Surface (MESS).

Main Methods:

• Conducted a soil survey analyzing 74 soil profiles in a 1,200-hectare steep-slope watershed.
• Applied MLR, C5-DT, and RF models, incorporating additional sampling to address dataset imbalance.
• Assessed accuracy using 5-fold cross-validation and extrapolation reliability with MESS analysis.

Main Results:

• The Random Forest model with additional sampling (RF*) demonstrated superior performance, achieving 49% overall accuracy and a 0.33 kappa index.
• RF* effectively linked soil mapping units (SMU) and identified occurrence conditions for less common soil classes within terrain attribute space.
• MESS analysis confirmed reliable model outputs for 82.5% of the study watershed.

Conclusions:

• The Random Forest model, enhanced with targeted additional sampling, is highly effective for digital soil mapping in challenging steep-slope environments.
• Integrating photointerpretation-derived sampling significantly improves the accuracy and reliability of spatial soil predictions.
• The study provides a robust methodology for soil mapping and highlights the importance of appropriate model selection and data augmentation.