This document provides an overview on the workflow that we used to
generate the models for predicting forest conservation value in
Denmark.
In brief:
- We gathered raster predictors with 10 m res. that we deemed
meaningful for predicting the conservation value of forests in
Denmark.
- We gathered ~20k annotations for forests with high and low
conservation value in Denmark.
- We generated a training dataset of 60k pixels that fell within the
annotated forests.
- We split the training dataset 80%/20% prior model training using a
geographic stratification.
- We trained Gradient Boosting and Random Forest models.
- We tuned the model hyperparameters using 5 or 10-fold cross
validation based on the training dataset from the 80%/20% split.
- We tested the final model performance on the validation dataset from
the 80%/20% split.
- We projected the forest quality across the whole of Denmark using
the final models and the predictor rasters.
