Open cnakalembe opened 12 months ago
@adebowaledaniel will be exporting the data next
✅ intercomparison results computed next steps: 1) update CEO sample notebook to allow stratification by classification ( @MsPixels will link here the GEE script used for this before), 2) add notebook to create ensemble or other map to visualize in GEE app ( @ivanzvonkov will work on 2 #346 )
note 9/11/23: will make the land cover stratification sample separate notebook from the existing CEO sample, under the new maps folder
Here's a GEE script to take a look at the ensembled map (worldcover, glad, esri): https://code.earthengine.google.com/1a3f781cca5f0d779ff4cec47afdd035
F1 score is 0.67 +/- 0.13, which is under our 0.7 threshold, so I'll train a model to see if we can obtain a better result.
Full intercomparison report: https://github.com/nasaharvest/crop-mask/blob/master/maps/Senegal_2022/intercomparison.ipynb
Trained model: https://github.com/nasaharvest/crop-mask/pull/347 Results:
"test_metrics": {
"accuracy": 0.8982,
"f1_score": 0.4615,
"precision_score": 0.5333,
"recall_score": 0.4068,
},
"val_metrics": {
"accuracy": 0.9612,
"f1_score": 0.697,
"precision_score": 0.7188,
"recall_score": 0.6765,
}
Not directly comparable to intercomparison results since the evaluation set is split. The test set f1 is low and the standard deviation is probably over 0.2, so thus far ensemle-subset map is my recommendation.
Next step: sign off from @cnakalembe if it looks good enough!
@cnakalembe will make a qualitative check-list for CSE to evaluate the map
re: checklist
re: map
Start year: 2022 Start month: April