MultiTemporal Crop Classification Example shown a really bad result.

[realleewq]

The MultiTemporalCropClassification experiment’s acc and f1-score is only 0.31, which is a really bad result. The graphical predict mask seem to be completely inconsistent with GT,
What went wrong with this experiment?

[romeokienzler]

@PedroConrado

[PedroConrado]

Thanks @realleewq for the feedback. The issue was due to an incorrect mode="min" in the ModelCheckpoint, which was corrected to mode="max". After updating the experiment, we achieved an f1-score of ~0.62. While this is a notable improvement, the results, and particularly the predict mask, could likely improve further with more than 30 epochs, as time constraints limited the training duration.

[tonyboston-au]

@PedroConrado Could the images in the example_multitemporalcrop.ipynb please be updated to match results obtained when using the mode="max" ModelCheckpoint? Current segmentation results are so bad that it doesn't inspire confidence in the utility of the Prithvi-EO-2.0 foundation model.

[romeokienzler]

@PedroConrado

[daniszw]

The notebook has been updated and now shows the correct predictions.

[tonyboston-au]

Thanks @daniszw