Run all commands from the tree_mapping directory.

Create and activate a conda environment using

conda env create -f tree_mapping.yml
conda activate tree_mapping

You will need to modify the path to the conda environment in all the bash files depending on where you store the environment.

First, you will need to get the 1m-resolution LiDAR-derived CHM files (from Esther or Jenia right now).

For comparison to existing data products, you will need to download:

• existing tree canopy height predicted maps from Lang et al. 2023, by downloading the relevant tiles from their tile browser
• existing tree canopy height maps from the GLAD project page
• existing tree canopy height maps from the WRI and Meta (hosted on AWS) -- or use acquire_data/download_meta_tiles.py
• For the Pauls map, run the download_pauls_map() function in utils.py. Note this requires creating a GEE project and typing the name into the ee.Initialize() function. Copy the tiff into tree_mapping/data/raw/global_tch_maps as pauls_map.tif.

Create a config file within the tree_mapping code directory called project_dir.yaml. Add a single line of the form project_dir: '../../../tambe_lab/Users/luciagordon/tree_mapping' that provides the relative path from the code directory to the directory where you want to store data, models, results, etc. This directory should be in a location with ~50GB available.

Second, from the code directory, run

bash run_data_preprocessing.sh

which will produce the data needed for training and evaluating our models:

1. Download the Sentinel-2 tiles that cover our sites
2. Aggregate LiDAR-derived CHM data by site (at native 1m resolution and coarsened 10m resolution)
3. Extract the per-site data from the existing global tree canopy height maps
4. Aggregate Sentinel-2 data and project to match the coarsened CHM data for each site

The notebook tree_mapping/process_data/visualize_data_preparation will show you what each of these steps is doing.

To generate the data for the ecological features against which to stratify performance, you will also need to obtain river_shapefile and geology_soil_vegetation_shapefile. Then you can run

python rasterize_line.py
python rasterize_polygon.py

to convert the river and geology shapefiles into tiffs, respectively. Then you can run

python feature_by_site.py

to extract the feature data for all the sites.

To compare to the globally trained models to predict tree canopy height from Lang et al. 2023, download their source code from the global-canopy-height-model GitHub page (v1.0), and put it at the same directory level as this repository. The configuration files for training with pretrained and randomly initialized models from that paper are in the experiment_configs folder.

First, you need to perform hyperparameter tuning for all the model settings with

bash run_hyperparameter_tuning.sh

Then run

bash run_10_seeds.sh

to train each model setting using 10 different random seeds and then use those trained models to generate canopy height predictions for all the sites. The train-test-validation splits are designed such that each site appears in the test set for precisely one split. To train models on a subset of the local data, use

sbatch run_train_nn_with_subsetted_train_sites.sh train_baseline_local_models.yaml channels 12 subset_train

where the config file, channels, and 12 will change depending on what model you want to train.

To evaluate local models and visualize predictions, run through the steps in eval_all_models.ipynb.

To train and produce maps for the the random forest models, use

python run_through_trained_RF_models.py

Run

python evalute_rf_output.py

to evaluate and visualize the output. The hyperparameter tuning results are generated by running

python hp_search_random_forest.py experiment_configs/tune_rf_12_channel.py

(or a different configuration file). To train models on a subset of the local data, use

python run_through_trained_RF_models.py

Run

sbatch run_tabulate_results.sh

in order to generate a JSON file containing all of the performance metrics needed to create the plots.

Notebooks to generate figures and tables are given in the figure_notebooks directory.