This set of changes is primarily focused on replacing Earth Engine with Descartes for accessing Sentinel data in the core pipeline stages. Listing key changes.
Pixel Classifier:
create_pixel_dataset: DL data pipeline. Operates on polygons, meaning that pixels outside of boundary are masked. Be aware that cloud filtering comes from DL sentinel cloud mask rather than S2 cloudless.
train_pixel_classifier: Some convenience changes and prediction simplification. Models now save with a .txt file listing a classification report and input data.
run_pixel_classifier: Repackaged to use the DL-backed version of nn_predict.py
nn_predict.py: Renamed the GEE version to nn_predict_gee.py. Replaced with a simpler and faster pixel classifier prediction script. Only key change is to specify an end_date rather than num_months as an argument.
Patch Classifier:
create_patch_dataset: DL data pipeline. Patches are now padded or trimmed to be a uniform shape. This was not guaranteed previously.
run_patch_classifier: DL data pipeline. Predictions are no longer made for only a single median patch. Instead, every patch below a threshold of cloudiness is evaluated. All predictions and patches are then stored in a dictionary.
Other
animate_patch: DL data pipeline.
descartes_model_run: Kick off a distributed model evaluation job on Descartes. Once finished, download the files and tile them into a single geotiff.
New sampling sites, including TPA site polygons for confirmed Java dumps.
New models. v1.1.8 is probably the current best?
@eboyda I think you're mostly familiar with these changes since you've been working off of the descartes-data branch. Mostly writing these notes to log what has changed.
This set of changes is primarily focused on replacing Earth Engine with Descartes for accessing Sentinel data in the core pipeline stages. Listing key changes.
Pixel Classifier:
create_pixel_dataset
: DL data pipeline. Operates on polygons, meaning that pixels outside of boundary are masked. Be aware that cloud filtering comes from DL sentinel cloud mask rather than S2 cloudless.train_pixel_classifier
: Some convenience changes and prediction simplification. Models now save with a .txt file listing a classification report and input data.run_pixel_classifier
: Repackaged to use the DL-backed version ofnn_predict.py
nn_predict.py
: Renamed the GEE version tonn_predict_gee.py
. Replaced with a simpler and faster pixel classifier prediction script. Only key change is to specify anend_date
rather thannum_months
as an argument.Patch Classifier:
create_patch_dataset
: DL data pipeline. Patches are now padded or trimmed to be a uniform shape. This was not guaranteed previously.run_patch_classifier
: DL data pipeline. Predictions are no longer made for only a single median patch. Instead, every patch below a threshold of cloudiness is evaluated. All predictions and patches are then stored in a dictionary.Other
animate_patch
: DL data pipeline.descartes_model_run
: Kick off a distributed model evaluation job on Descartes. Once finished, download the files and tile them into a single geotiff.@eboyda I think you're mostly familiar with these changes since you've been working off of the
descartes-data
branch. Mostly writing these notes to log what has changed.