This contains code for the training the forest loss driver model, as well as visualizing the outputs of the prediction pipeline.

The prediction pipeline itself has been integrated into rslp and some of this code should be moved there later too.

Training and Prediction

Train the model:

python -m rslp.rslearn_main model fit --config amazon_conservation/train/config_satlaspretrain_flip_oldmodel_unfreeze.yaml

Run the prediction data preparation pipeline, this populates an rslearn dataset with windows corresponding to recent forest loss events:

python -m rslp.main forest_loss_driver predict --pred_config.workers 32 --pred_config.ds_root gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/

Replace YYYYMMDD with the date you are starting to run the pipeline on. You can reduce the time range to predict using --pred_config.days X (the default is 365 to predict over the previous year).

Now obtain the images, this will take many hours:

rslearn dataset prepare --root gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/ --workers 128 --batch-size 4
rslearn dataset ingest --root gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/ --workers 64 --jobs-per-process 1 --no-use-initial-job
rslearn dataset materialize --root gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/ --workers 64 --no-use-initial-job --batch-size 4

Run the second step of the pipeline which identifies the least cloudy images, these are the ones that the model will input:

python -m rslp.main forest_loss_driver select_best_images --ds_path gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/

Now the model can be applied:

python -m rslp.rslearn_main model predict --config data/forest_loss_driver/config_satlaspretrain_flip_oldmodel_unfreeze.yaml --data.init_args.path gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/

Visualization

After the prediction pipeline completes, we can visualize the forest loss events together with the predicted drivers. First create a JSON file containing the windows that have outputs. Some windows will not have had enough non-cloudy Sentinel-2 images to compute outputs.

python -m amazon_conservation.predict.find_windows_with_outputs --ds_path gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/ --out_fname good_windows.json

Collect all of the forest loss events with the predicted driver into a GeoJSON file:

python -m amazon_conservation.predict.collect_outputs_to_geojson --ds_path gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/ --windows_fname good_windows.json --out_fname forest_loss_events.geojson

Use tippecanoe (requires installation from source) to create vector tiles of the GeoJSON suitable for loading in the web app:

tippecanoe -zg -e out_tiles/ --drop-smallest-as-needed --no-tile-compression forest_loss_events.geojson
# Copy to a public bucket, e.g.:
aws s3 --endpoint-url https://b2bcf985082a37eaf385c532ee37928d.r2.cloudflarestorage.com sync out_tiles/ s3://satlas-explorer-data/rslearn-public/forest_loss_driver/YYYYMMDD/tiles/

You will need to change the let pbfLayer = L.vectorGrid.protobuf in map.html to match the path where the tiles can be read by web browser. Now launch the webserver:

cd amazon_conservation/predict
python server.py gs://rslearn-eai/datasets/forest_loss_driver/prediction/dataset_YYYYMMDD/ 8080 ../../good_windows.json