Cropland: Tanzania North 2019

[ivanzvonkov]

Key tasks:

• [x] Verify the quality of Tanzania labels (set 1, set 2) on CEO
• [x] Train a model using Tanzania data for evaluation
• [x] Use Tanzania model to create a crop mask for North Tanzania
• [x] Comments about quality of map from @cnakalembe

[adebowaledaniel]

Completed! View in EE here

[ivanzvonkov]

Reposting slack message here: Looks pretty good upon visual inspection

• Water bodies are classified as non-crop
• The crop % seems to be around half which is close to recent estimates
• Some cropland is missed on the East
• Kilimanjaro is correctly classified as non-crop 🏔️

Looks like many of those missing values you got were in the ocean I also see a decent amount of ERA5 looking grid artifacts, so next step would be to retrain the model with the skip_ERA5 argument and rerun predictions. This time the data is all already exported so no need to wait for Earth Engine.

[adebowaledaniel]

Pull request for the skip_ERA5 model (https://github.com/nasaharvest/crop-mask/pull/205).

[adebowaledaniel]

Tanzania North 2019 (skip ERA5) crop mask view here.

[ivanzvonkov]

Tanzania North 2019 (skip ERA5) crop mask view here.

The script has an issue: Line 1: image2 is not defined

Could you also add both results to the same script so we can compare? @adebowaledaniel

[adebowaledaniel]

Sorry about that; here is the updated link.

With both results, you mean with and without ERA5? @ivanzvonkov

[ivanzvonkov]

Yes both with and without ERA5

[adebowaledaniel]

You can compare both results here

[adebowaledaniel]

@ivanzvonkov The updated comparison view is here. I will mosaic the multiple rasters generated for each model if needed in another issue.

[ivanzvonkov]

Excellent! Looks like there are some more fields found in the North East now and a lot fewer grid artifacts. I didn't realize that the f1-score is also higher when skipping ERA5! 🤔 https://github.com/nasaharvest/crop-mask/blob/494d1d1d7e21dff64e2fcafe2e12887f6d680e7d/data/models.json#L87

@gabrieltseng @hannah-rae interesting result here!

[cnakalembe]

Reviewed and ready to merge with Tanzania South

[adebowaledaniel]

Excellent! Looks like there are some more fields found in the North East now and a lot fewer grid artifacts. I didn't realize that the f1-score is also higher when skipping ERA5! 🤔

https://github.com/nasaharvest/crop-mask/blob/494d1d1d7e21dff64e2fcafe2e12887f6d680e7d/data/models.json#L87

@gabrieltseng @hannah-rae interesting result here!

@ivanzvonkov, what is the following line of action with this? Is there a conclusion on the performance of with/without ERA5 based on other models trained? Moving forward, which model versions should be used for the final merge of both Tanzania North and South?