nasaharvest / crop-mask

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

Cropland: Sudan Blue Nile 2019 #217

Closed ivanzvonkov closed 1 year ago

ivanzvonkov commented 1 year ago

Start year: 2019 Start month: February

ivanzvonkov commented 1 year ago

Data being added to repo: https://github.com/nasaharvest/crop-mask/pull/230

ivanzvonkov commented 1 year ago

Data stats

https://github.com/nasaharvest/crop-mask/blob/0dc58b8867dbee8fa17fa5131c95e52e97d3fcb3/data/report.txt#L269

Sudan_Blue_Nile_CEO_2019 (Timesteps: 24)
----------------------------------------------------------------------------
disagreement: 0.0%
eo_data_complete    1500
✔ training amount: 311, positive class: 31.5%
✔ validation amount: 581, positive class: 37.9%
✔ testing amount: 608, positive class: 37.5%

Test metrics

https://github.com/nasaharvest/crop-mask/blob/0dc58b8867dbee8fa17fa5131c95e52e97d3fcb3/data/models.json#L70

"accuracy": 0.9408,
"f1_score": 0.9227,
"precision_score": 0.9034,
"recall_score": 0.943,
"roc_auc_score": 0.9723

Map

https://code.earthengine.google.com/382a613f09d8d435ee0ccd9313bde813

image

Additional Notes

Initial prediction merging led to some missing tiles, after investigating and remerging same predictions v2 had not missing tiles. Attributing to some intermittent merging issue.

ivanzvonkov commented 1 year ago

Related #244

cnakalembe commented 1 year ago

Reviewed-- as with others like Tanzania, we seem to have issues around wetlands; otherwise, the mask looks decent. I suggest we look into sourcing non-crop examples of wetlands in CropHarvest possibly from a recent wetland product-- hopefully, one that does well at separating crops like rice.

hannah-rae commented 1 year ago

Potential source of global wetland map that we could use, but should cross-check with the "problem areas" where wetlands are misclassified in our map: https://www2.cifor.org/global-wetlands/

hannah-rae commented 1 year ago

Another potential wetlands dataset for sampling non-crop points from: https://www.worldwildlife.org/pages/global-lakes-and-wetlands-database

ivanzvonkov commented 1 year ago

Next step corrective labeling with GEE script, relevant #222

hannah-rae commented 1 year ago

Options in Earth Engine: https://developers.google.com/s/results/earth-engine/datasets?q=wetlands&text=wetlands

ivanzvonkov commented 1 year ago

Short term -> corrective labeling Long term -> Issue about wetland data + analysis

MsPixels commented 1 year ago

@cnakalembe, please find the link to the Corrective labeling App for Sudan Blue Nile https://aasareansah.users.earthengine.app/view/sudanbluenilecorrectivelabelapp

cnakalembe commented 1 year ago

Sudan Blue Nile_new points.csv

ivanzvonkov commented 1 year ago

Second version of map: https://code.earthengine.google.com/1026aec46650a3578d3672b04e34c4d2 Looks about the same

cnakalembe commented 1 year ago

It is better in some places. I can do another round of corrective labeling, this time denser than before. The first time I was being conservative, looking for larger patterns.

ivanzvonkov commented 1 year ago

Let's discuss on Monday in the meeting!