Week 14: Spatial Machine Learning and Model Evaluation¶

Core Questions¶

• Why can random train/test splits exaggerate ecological model performance?
• How does spatial leakage occur?
• How should transferability, extrapolation, and uncertainty be evaluated?

Concepts¶

• Train/test splitting.
• Spatial cross-validation.
• Spatial leakage.
• Overfitting.
• Model transferability.
• Environmental extrapolation and novelty.
• Model uncertainty.
• Variable importance and partial dependence.

Python Tools¶

• scikit-learn.
• GeoPandas.
• NumPy.

Applied Lab¶

Students compare a random train/test split with a spatial train/test split and evaluate how the estimated predictive performance changes.

Deliverable¶

Submit a model evaluation report with cross-validation design, performance comparison, extrapolation or uncertainty discussion, and a justified final modeling approach.