nasaharvest / crop-mask

End-to-end workflow for generating high resolution cropland maps
Apache License 2.0
95 stars 28 forks source link

Cropland: Tigray Ethiopia 2020 #252

Closed hannah-rae closed 1 year ago

hannah-rae commented 1 year ago

Start year: 2020 Start month: February

hannah-rae commented 1 year ago

The labels for validation/test for this have already been completed last year. The data associated with the labels has been exported too, but I think we should re-export them since there have been changes since the last time it was done. @ivanzvonkov do you agree?

ivanzvonkov commented 1 year ago

Yes you can see the data that is already available here: https://github.com/nasaharvest/crop-mask/blob/74913ceff6b4bc4c3fd88f0aa3a381043108ffc9/data/report.txt#L132

The csv files can be accessed by running the following inside a cloned repo

dvc pull data/datasets

The data has been updated according to various changes over the last year. One outstanding change is using data with nans instead of skipping it. This can be achieved by First, deleting the csv from data/datasets Then

git checkout -b'updating-tigray-data'
dvc commit data/datasets 
dvc push
git add .
git commit -m'Update data'
git push

This will run the data pipeline and automatically generate new csvs with missing data not skipped.

hannah-rae commented 1 year ago

Model is trained, pending merge of #266

bhyeh commented 1 year ago

map preview

  1. Missing data - though create_map.ipynb does not indicate that there is more available data.
bhyeh commented 1 year ago

Retrain model without ERA5 data. Middle artifacts present in above seem to be in line with ERA5 sized tiling - Dr. Kerner

bhyeh commented 1 year ago

Retrained 2020 model without ERA5 data. Changed testing and validation metrics here

ivanzvonkov commented 1 year ago

@bhyeh is the new map ready for review? Does it have artifacts and/or missing data?

bhyeh commented 1 year ago

Here is a link for the latest map for 2020. Was having issues with missing predictions the last few tries - but all is well now!

hannah-rae commented 1 year ago

Awesome thanks @bhyeh! Looking at the map, I think it seems a bit worse than the 2021 one, which is also shown in the metrics. I am wondering if we can boost the performance a bit with corrective labeling.

@MsPixels Would you be able to help @bhyeh get a corrective labeling project set up for this map?

MsPixels commented 1 year ago

Yes @hannah-rae, I can set it up for @bhyeh

MsPixels commented 1 year ago

@bhyeh, Corrective labeling for Tigray 2020. More resources can also be found here

hannah-rae commented 1 year ago

Next step: @bhyeh to make updated Tigray 2020 map with model updated with corrective labels