Spatial Decision Support System

For farmers

This tool estimates the organic matter in your soil (SOM = Soil Organic Matter) using satellite images. The model was trained on Quebec farm measurements and links satellite-observed bare-soil colour signals to lab SOM values.

R² (R-squared) measures how well the model fits — 1.0 is perfect. Above 0.75: excellent. 0.50–0.75: good. Below 0.25: use with caution. It tells you how confident the model is in its prediction for this field.

RMSE and MAE show the average prediction error in the same units as SOM (g/kg or %). Lower is better. An RMSE below 3 g/kg (0.3%) is generally acceptable for field-scale SOM.

Images used: the number of cloud-free satellite images from autumn and early spring included in the model. More images = more reliable result. At least 5 is recommended.

For scientists

RandomForestRegressor on Sentinel-2 spectral indices (BSI, NDVI, EVI, SAVI, OMI, CAI, RI, SI, BI, CI, NDMI + stdDev) with one-hot soil type encoding. Scenarios: S1 (spectral+soil), S2 (+topo), S3 (+topo+climate). Feature selection via LassoCV on inner train set (GroupShuffleSplit 80/20 by FIELD_ID). Image ranking via GroupKFold(3) RMSE. Optimal image count: argmax R² on inner validation. Back-transform: Duan smearing (log₁₀).

R² and RMSE are computed on field-level aggregated predictions (mean over images per field). When the test set contains only 1 field, field-level R² is undefined; inner-validation R² is shown instead (*).

Training data from GEE: Sentinel-2 L2A, bare-soil masked (BSI > 0), cloud filter < 1% at tile level. Low image counts may reflect strict cloud filtering in high-cloud years or limited bare-soil exposure windows.