Week 07: Building a complete SDM w/ a machine learning model

Icebreaker

If you could bring back any fashion trend, what would it be?

Paper discussion

• Allie will be leading the discussion this week of Aoyama et al 2020 "Incorporating diversity measures into Ecological Site Descriptions to manage biodiversity on heterogeneous landscapes"

Lab Exercise: Building a more complete SDM w/ Machine Learning

Core Questions

• How does this 'machine learning' stuff actually work?
• How does a random forest learn a species–environment relationship?
• How does prediction error change when we add more environmental variables?
• Which bioclim variables does the ML model use most, and does importance imply an ecological cause?

Concepts

• Features vs Targets (machine learning jargon for predictor vs response)
• Decision Trees
• Train/test data splitting
• Comparing one-variable and multiple-variable models
• Prediction evaluation; Mean absolute error (MAE)
• Variable importance

Applied Lab

• A visual introduction to machine learning SDMs

• Next, the usual intro update exercise: Go to your JupyterHub and pull the latest version of the class github repository.

Python Tools

• scikit-learn
• GeoPandas
• NumPy
• pandas
• xarray
• Matplotlib

Students build a presence/background SDM for a Maine amphibian using scikit-learn's RandomForestRegressor and the supplied occurrence–climate table and Maine bioclimatic raster stack. In short, guided steps, they first train a forest using annual mean temperature (bio_1) and evaluate MAE on withheld cells. They add annual precipitation (bio_12) and precipitation seasonality (bio_15), refit using the same training cells, and compare MAE on the same test cells. They then map suitability and plot variable importance for the three-variable model.

• Lab 07: Machine Learning SDMs

Assignment

Repeat the workflow for a chosen amphibian species, justify predictor roles, and compare the BIO1-only and three-variable models on the same test cells. Submit a completed notebook with maps, short interpretations, and one saved suitability raster. Use the lab's model settings throughout.

• Read A visual introduction to machine learning (It's short and fun).
• Assignment 07: Machine Learning SDMs

Paper discussion leader next week: Next week we will need to do 2 papers so everyone has a chance to lead over the semester. So, Meg & Hao! Please each select a (hopefully shorter) paper for the group to discuss before Sunday 10/18 6pm and send it around to the class email list: bio597-fall2026-group@maine.edu