Global soil products such as SoilGrids provide invaluable information, but their spatial resolution (≈250 m) is often too coarse for local agricultural decision-making. Farmers and land managers operate at much finer spatial scales.
This project demonstrates an open-science, reproducible workflow to generate high-resolution (30 m) Soil Organic Carbon (SOC) maps by downscaling coarse global soil data using high-resolution satellite imagery and terrain information.
The workflow is implemented as a Google Colab notebook and is designed for:
- teaching and training
- rapid prototyping
- transparent, cloud-native geospatial analysis
- Access cloud-hosted geospatial data (Landsat, NASADEM, SoilGrids)
- Engineer environmental predictors relevant to soil processes
- Downscale coarse SOC data using machine learning
- Produce a fine-resolution SOC map with uncertainty estimates
- Demonstrate open, reproducible geospatial science
- Region: Tell Atlas, Northern Algeria
- Extent: ~10 × 10 km agricultural window
- Spatial Resolution: 30 m (output)
-
Target Variable (Teacher)
- SoilGrids SOC stock (0–30 cm depth, ~250 m)
-
Predictors (Students)
- Vegetation indices from Landsat:
- MSAVI (Modified Soil-Adjusted Vegetation Index)
- Bare Soil Index (BSI)
- Topographic variables:
- Elevation (NASADEM)
- Slope
- Vegetation indices from Landsat:
-
Downscaling Strategy
- Resample coarse SOC to the 30 m grid
- Learn relationships between SOC and environmental predictors
- Predict SOC continuously at high resolution
-
Model
- Random Forest Regressor (scikit-learn)
-
Uncertainty Quantification
- Ensemble approach (multiple RF models)
- Pixel-wise standard deviation as uncertainty proxy
The full workflow is implemented in the notebook: https://github.com/OpenClimateScience/Module_4_Open_Science_for_Soil_Health_FR/blob/main/Module_4_Open_Science_for_Soil_Health_FR.ipynb