How Predictions Work

How predictions work

Canopi predicts the probability that a tree planted at a specific location, of a specific species, using a specific planting method, will be alive at 1 year, 3 years, and 5 years after planting.

What a survival probability means

When Canopi returns survival_probability: 0.82 at a 5-year horizon, it means: based on historical patterns of tree survival across thousands of sites with similar soil, climate, terrain, and planting conditions, approximately 82% of trees are expected to survive to year five at this location.

This is a statistical probability, not a guarantee. Forests are complex systems influenced by weather extremes, pest outbreaks, fire, and human activity that no model can predict with certainty. A probability is honest about that uncertainty.

How the prediction is generated

1. You send coordinates, a species, and a planting method. These define the prediction question: “What happens if I plant this species here, this way?”

2. Canopi finds the nearest data point. Our prediction network covers approximately 18,000 sites across Oregon and Washington. The API locates the closest site with pre-computed predictions and reports the match distance so you know how close the data is to your actual location.

3. Pre-computed predictions are returned. Canopi’s XGBoost model has already evaluated every site-species-method-horizon combination. The predictions account for 17 environmental features including soil composition, climate patterns, terrain, and the planting method itself.

4. Risk factors are extracted. Each prediction includes the top environmental factors influencing survival at that site, derived from SHAP (SHapley Additive exPlanations) analysis. These aren’t generic — they’re specific to this site, this species, and this horizon.

The features behind the prediction

The model evaluates 17 features for every prediction:

Soil characteristics — Organic matter content, available water capacity, clay percentage, and drainage class. These determine how well a site retains and delivers moisture to roots.

Climate patterns — Precipitation, maximum temperature, mean temperature, dew point temperature, and maximum vapor pressure deficit. These capture the atmospheric conditions that drive water stress — the primary killer of seedlings.

Terrain — Elevation. Combined with climate features, this captures the temperature and moisture gradients that define where species can thrive.

Tree characteristics — Crown ratio, height, and diameter at the baseline measurement. These reflect the condition and vigor of trees at similar sites in the training data.

Planting method — Manual versus drone-seeded. The model learns how method choice interacts with site conditions to affect survival outcomes.

Time horizons

Canopi currently predicts across three horizons:

• 1 year — First-year establishment. This is when the highest mortality occurs. Seedlings face transplant shock, drought stress, frost, and competition from established vegetation. First-year survival is the most actionable prediction — it’s the horizon where intervention (irrigation, weed control, replanting) is most feasible.

• 3 years — Mid-term survival. Trees that survive the first year face ongoing climate stress and competition. The 3-year mark is a common milestone for reforestation success assessment.

• 5 years — Establishment threshold. In most reforestation programs and carbon credit methodologies, trees surviving to year five are considered established. This is the horizon most relevant to financial decisions — carbon credit forward contracts, project insurance, and investment underwriting.

Match distance

Every response includes match_distance_km — the distance between your requested coordinates and the nearest data point Canopi has predictions for. This is a transparency feature.

• Under 2km — Very close match. Predictions are highly site-relevant.
• 2-5km — Good match. Environmental conditions are likely similar.
• 5-10km — Moderate match. Terrain and microclimate may differ from the matched site.
• Over 10km — Weak match. Use predictions directionally rather than precisely.