Dataset for: “From Biological to Structural Controls: Explainable Machine Learning Reveals a Fundamental Shift in Soil Carbon Drivers Across Depth in Subtropical Plantations”
Description
This dataset supports the findings of the manuscript entitled “From Biological to Structural Controls: Explainable Machine Learning Reveals a Fundamental Shift in Soil Carbon Drivers Across Depth in Subtropical Plantations,” submitted to the journal “Catena”. Study Overview: The study aimed to quantitatively uncover the depth-dependent drivers of soil organic carbon (SOC) and microbial biomass (MBC, MBN) in subtropical plantations using explainable machine learning (Random Forest and XGBoost coupled with SHAP analysis). Data was collected from 13-year-old monoculture and intercropping plantations of three species (Ligustrum lucidum, Osmanthus fragrans, and Cinnamomum camphora) in the Shihe River Basin, Henan Province, China. Data Content: The dataset comprises two main parts: 1. Stand Inventory Data: Tree species, planting pattern (monoculture/intercropping), mean diameter at breast height (DBH, cm), mean tree height (m), canopy density (%), and stand density (trees/ha). 2. Soil Property Data: Samples were collected from two depth intervals (0-10 cm and 30-50 cm). Analyzed properties include: Soil organic carbon (SOC, g/kg), Particulate organic carbon (POC, g/kg), Easily oxidized organic carbon (EOC, g/kg), Total nitrogen (TN, mg/kg), Ammonium nitrogen (NH₄⁺-N, mg/kg), Microbial biomass carbon (MBC, mg/kg), Microbial biomass nitrogen (MBN, mg/kg). Data Availability and Usage: This dataset is embargoed until December 31, 2026 to allow for the publication of the associated manuscript. Upon expiration of the embargo, the data will be fully accessible under a CC-BY 4.0 license. Researchers are encouraged to cite this dataset if used in their work.
Files
Steps to reproduce
The following steps outline how to reproduce the primary machine learning modeling and analysis presented in the associated manuscript using the provided data file “data.xlsx”. 1. Data Preprocessing: Load the dataset. Encode categorical variables (e.g., “Tree Species”, “Planting Pattern”) using one-hot encoding or label encoding. Split the dataset into a training set (70%) and an independent test set (30%), ensuring stratification by “Soil Depth”. 2. Model Training: Implement the two ensemble algorithms as described: Random Forest (RF): Using the scikit-learn library in Python, train a MultiOutputRegressor with a RandomForestRegressor base estimator. Use the hyperparameters: n_estimators=200, max_depth=10. XGBoost: Using the xgboost library, train a MultiOutputRegressor with an XGBRegressor base estimator. Use the hyperparameters: n_estimators=150, max_depth=6, learning_rate=0.1. 3. Model Evaluation: Predict the three targets (SOC, MBC, MBN) on the independent test set. Calculate performance metrics (R², RMSE, MAE) to compare with the results in Table 2 of the manuscript. 4. SHAP Analysis (Core Interpretation): For the trained models (recommend using the best-performing one for each target), calculate SHAP values using the Python `shap` library (e.g., “shap.TreeExplainer”). Plot the global feature importance (mean absolute SHAP value) to reproduce figures like Fig. 3. Generate SHAP dependence plots for key features like “DBH” and “Stand Density” to reveal the non-linear and threshold effects shown in Fig. 4b and 4c. Calculate and visualize SHAP interaction values to confirm the synergistic effect between 'C. camphora' and 'Intercropping' (Fig. 4d). 5. Statistical Summary: Basic descriptive statistics (mean, standard deviation) grouped by 'Soil Depth' can be computed to reproduce the trends shown in Table 1.
Institutions
- Xinxiang University