Tutorial 1: How predictive maps are made

From field measurements and satellite data to continuous maps of land health indicators.

1.1 What is a predictive map?

• A predictive map estimates a variable (e.g., SOC, erosion risk, vegetation condition) at locations where we have no direct measurement.
• It is built by learning relationships between:
  • What we observe in the field (soil samples, vegetation surveys), and
    
  • Other spatial layers (rainfall, topography, land cover, etc.) that cover the whole region.
• The result is a continuous surface where every pixel has a predicted value based on those relationships.
• Predictive maps are always estimates, not a perfect representation of reality.

1.2 Inputs: field data, satellites, and covariates

• Field data (observations):
  • Measurements at specific locations (e.g., SOC from LDSF plots, tree counts, erosion assessments).
  • These points are our “ground truth” used to train and check the model.
• Satellite and other spatial data (covariates):
  • Climate: rainfall, temperature, drought indices.
  • Land: elevation, slope, landform, soils, land cover.
  • Vegetation: EVI, NDVI, tree cover, fire history.
• In a predictive model:
  • Field measurements are the target (what we want to predict).
  • Satellite and other layers are predictors (variables that explain how the target varies in space).

1.3 Building the model

• Step 1: Align data
  • Extract predictor values (e.g., rainfall, EVI, elevation) at each field measurement location.
  • Build a table where each row is a field sample and each column is a predictor.
• Step 2: Fit a model
  • Use statistical or machine-learning methods (e.g., regression, random forests, gradient boosting) to learn:
    • How changes in predictors (rainfall, slope, vegetation) are associated with changes in the target (SOC, erosion risk, etc.).
• Step 3: Predict across the landscape
  • Apply the trained model to every pixel in the region using the predictor layers.
  • For each pixel, the model estimates the likely value of the target indicator.
  • The output is a raster where each pixel holds the predicted value.

1.4 Validation and quality checks

• Good predictive maps must be tested, not just produced.
• Common checks include:
  • Holding back some field data as a test set that is not used for model training.
  • Comparing model predictions against these held-out observations.
  • Calculating summary statistics such as:
    • How close predictions are, on average, to observed values.
    • Whether the model underestimates or overestimates in certain environments.
• Additional checks:
  • Looking at maps of residuals (differences between predicted and observed values).
  • Evaluating performance across different subregions or land types.
• These tests help identify:
  • Where the model is strong, and
    
  • Where we should be more cautious in interpreting the map.

1.5 From model output to decision-ready maps

• Raw model outputs often go through additional steps before reaching decision-makers:
  • Masking areas outside the data domain (e.g., water bodies, urban areas, non-target ecosystems).
  • Smoothing or aggregation to reduce noise and match the scale of decision-making (e.g., administrative units or watersheds).
  • Reclassification into categories (e.g., low/medium/high SOC, low/high erosion risk) to make interpretation easier.
• For K4GGWA, these processed maps are then:
  • Integrated into dashboards and reports.
  • Linked with contextual information (climate, land use, communities).
  • Used alongside uncertainty information (explained in Tutorial 3) to support responsible decisions.
• The key message:
  • Predictive maps are data-driven approximations built from field and satellite information.
  • Their value depends on both the quality of inputs and the rigour taken in modelling and validation.