Soil Organic Carbon (SOC) Prediction for Mato Grosso, Brazil

This project presents a spatial modeling framework for predicting Soil Organic Carbon (SOC) concentrations across Mato Grosso, Brazil, using machine learning techniques and Earth observation data. The study integrates field-sampled SOC data with environmental predictors extracted from Google Earth Engine (GEE) to build a Random Forest model using Python’s scikit-learn library. The model was trained over a large-scale area (~950,000 km²), representing one of the most agriculturally dynamic regions of the Brazilian Cerrado. Key environmental variables used as predictors included Land Surface Temperature (LST), rainfall, elevation, slope, Normalized Difference Vegetation Index (NDVI), Bare Soil Index (BSI), and land use. The feature importance analysis revealed that LST (31%), rainfall (28%), and elevation (26%) were the most influential in predicting SOC, emphasizing the dominant role of climate and terrain in SOC distribution. Variables like NDVI, BSI, and land use had lower contributions, indicating that vegetation and surface indices alone do not fully account for SOC variability across the landscape. The model operated at a spatial resolution suitable for regional-scale analysis. This approach demonstrates the effectiveness of combining remote sensing data from GEE and Python-based machine learning tools to inform soil health assessment and land management strategies. The results from this study have been published, confirming the robustness and scientific relevance of the methodology.

Study Area

Mato Grosso is a vast state in central-western Brazil and one of the most significant regions in the Brazilian Cerrado biome, known for its biodiversity and intense agricultural expansion. The area spans diverse land use patterns—from croplands and pastures to natural savannas—making it a vital landscape for studying carbon storage potential and land degradation. This work focuses on modeling SOC dynamics across the entire state (~950,000 km²).

Environmental Predictors

Key environmental variables used as predictors included:

• Land Surface Temperature (LST)
• Rainfall
• Elevation
• Slope
• Normalized Difference Vegetation Index (NDVI)
• Bare Soil Index (BSI)
• Land use

Workflow Architecture

The pipeline integrates geospatial data processing, field validation, and model-driven prediction.

Methodology

1. Field Data Collection

SOC values were collected from soil samples across representative sites in Mato Grosso. These values were geocoded and stored as point shapefiles (sampling_points.shp), covering various land use types.

2. Predictor Variable Extraction (Google Earth Engine)

Using Google Earth Engine (GEE), multi-temporal and environmental predictor variables were extracted for the study area:

• Land cover (from MODIS and ESA datasets)
• Topographic attributes (elevation and slope from SRTM)
• Climate data (monthly temperature and precipitation)
• Remote sensing indices: NDVI (Normalized Difference Vegetation Index), BSI (Bare Soil Index)

All rasters were resampled to a uniform spatial resolution and aligned to the study grid.

3. Modeling Approach

• Exploratory Data Analysis (EDA) and OLS regression were first conducted to evaluate the relationships between SOC and environmental variables.
• A Random Forest regression model was trained using 70% of the sample data and validated on the remaining 30%.
• Key metrics: R², RMSE, and MAE were computed to evaluate model performance.

4. Prediction

The trained model was applied to unsampled locations using the predictive_dataset_label.shp grid. Outputs were exported as:

• predicted_data.csv – SOC values for new locations
• soc_pred.geojson – Spatial map for visualization

Data Description

File	Description
sampling_points.shp	Georeferenced field SOC measurements
Matto_Grosso.shp	Administrative boundary of the study area
predictive_dataset_label.shp	Prediction grid with extracted variables for inference

Results

• The Random Forest model demonstrated high accuracy in predicting SOC distribution across the state.
• Feature importance analysis identified NDVI, slope, rainfall, and elevation as major contributors.
• The prediction provides a powerful visual and analytical tool for guiding land management and restoration strategies in the Cerrado.

Visual Outputs

Land Use Map

Terrain Map (Slope & Elevation)

Remote Sensing Indices (NDVI and BSI)

Climate Variables (Rainfall & LST)

Predicted SOC Map

Model Importance

Importance of the Study to Forestry and Agriculture/Food Security

• Improved Soil Health Assessment: Enables accurate evaluation of soil fertility crucial for sustainable forestry and agriculture.
• Climate and Terrain Influence Understanding: Reveals key environmental factors guiding tailored land management decisions.
• Support for Sustainable Land Use Planning: Helps prioritize areas for conservation, reforestation, and agricultural improvement.
• Enhanced Crop Yield and Forest Growth Predictions: Links SOC to nutrient and water availability for better productivity forecasts.
• Integration of Remote Sensing for Large-Scale Monitoring: Provides cost-effective, large-area soil condition monitoring to aid timely decisions.

Usage

1. Open the Jupyter Notebook soc_prediction_v2.ipynb.
2. Follow the steps from data loading, preprocessing, and exploratory analysis to model training and spatial prediction.
3. View and export results for further analysis or visualization in GIS platforms.

To clone this repository:

git clone https://github.com/sammygis/matto-grosso-soc-prediction.git

References

1. Nwaogu, C., Diagi, B.E., Ekweogu, C.V. et al. (2024). Soil organic carbon stocks as driven by land use in Mato Grosso State: the Brazilian Cerrado agricultural frontier. Discover Sustainability, 5, 382. https://doi.org/10.1007/s43621-024-00592-w

   This paper was published as a direct result of this research.

Acknowledgments

This work was inspired in part by Hydroinformatics: Urban Flood Susceptibility Mapping, which guided the integration of Earth observation and machine learning.