Skip to content

Latest commit

 

History

History

sentinel2_postprocess_revamp

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
README.md
README.md
 
 
config.json
config.json
 
 
convert_to_multisat.py
convert_to_multisat.py
 
 
create_windows.py
create_windows.py
 
 
get_label_csv.py
get_label_csv.py
 
 
rslearn_entrypoint.py
rslearn_entrypoint.py
 
 
ship_types.py
ship_types.py
 
 
train_config_ms.yaml
train_config_ms.yaml
 
 
train_config_rgb_crop_imagenet.yaml
train_config_rgb_crop_imagenet.yaml
 
 
train_config_rgb_flip_imagenet.yaml
train_config_rgb_flip_imagenet.yaml
 
 
train_config_rgb_flip_satlaspretrain.yaml
train_config_rgb_flip_satlaspretrain.yaml
 
 
train_config_rgb_flip_satlaspretrain_nonorm.yaml
train_config_rgb_flip_satlaspretrain_nonorm.yaml
 
 
train_config_rgb_flipcrop_imagenet.yaml
train_config_rgb_flipcrop_imagenet.yaml
 
 
train_config_rgb_noflip_imagenet.yaml
train_config_rgb_noflip_imagenet.yaml
 
 
train_config_rgb_noflip_satlaspretrain.yaml
train_config_rgb_noflip_satlaspretrain.yaml
 
 
train_config_rgb_noflip_satlaspretrain_fpn.yaml
train_config_rgb_noflip_satlaspretrain_fpn.yaml
 
 
train_config_rgb_noflip_satlaspretrain_freeze.yaml
train_config_rgb_noflip_satlaspretrain_freeze.yaml
 
 
visualize.py
visualize.py
 
 
visualize_html.py
visualize_html.py
 
 

README.md

Data

The input data for this project component are the CSV files that Hunter created containing Sentinel-2 vessel position + AIS metadata.

There are currently two formats:

  • First v1 CSV (sentinel2_vessel_labels_with_metadata.csv) with metadata for a subset of the vessel labels in the Sentinel-2 vessel detection dataset.

  • Later v2 CSVs with metadata for recent AIS-correlated vessel predictions.

We get image crops centered at vessels using rslearn. First create an empty dataset e.g. at /data/favyenb/rslearn_sentinel2_vessel_postprocess.

Copy config.json to the dataset directory.

Then create windows from the different CSVs:

python create_windows_v1.py /home/favyenb/sentinel2_vessel_labels_with_metadata.csv /data/favyenb/rslearn_sentinel2_vessel_postprocess labels
python create_windows_v2.py /home/favyenb/sentinel2_correlated_detections.csv /data/favyenb/rslearn_sentinel2_vessel_postprocess detections

Then use rslearn to prepare, ingest, and materialize the dataset:

python -m rslearn.main dataset prepare --root /data/favyenb/rslearn_sentinel2_vessel_postprocess/ --workers 64 --group detections --batch-size 8
python -m rslearn.main dataset ingest --root /data/favyenb/rslearn_sentinel2_vessel_postprocess/ --workers 64 --group detections --no-use-initial-job --jobs-per-process 1
python -m rslearn.main dataset materialize --root /data/favyenb/rslearn_sentinel2_vessel_postprocess/ --workers 64 --group detections --no-use-initial-job --jobs-per-process 1

The materialized dataset as of 2024-06-12 (17K vessels) is available here:

gs://satlas-explorer-data/rslearn_labels/rslearn_sentinel2_vessel_postprocess_2024-06-12.tar

Training

train_config.yaml provides the configuration to train the model with rslearn. The only custom component needed is a visualization function provided by rslearn_entrypoint.py.

Train the model:

PYTHONPATH=/path/to/rslearn python rslearn_entrypoint.py model fit --config ../rslearn_projects/sentinel2_postprocess_revamp/train_config.yaml
mkdir vis
PYTHONPATH=/path/to/rslearn python rslearn_entrypoint.py model test --config ../rslearn_projects/sentinel2_postprocess_revamp/train_config.yaml --ckpt_path lightning_logs/version_0/checkpoints/... --model.visualize_dir vis/