U-Net generally conforming to the current
FTW baseline 3-class model settings, trained on the Fields of the World full dataset, using both time points (Window A, Window B) for 100 epochs. The model is the same as the current FTW PRUE variant with EfficientNet-b7 backbone, but was trained using a locally-weighted Tversky Focal Loss rather Logcoshdice, and using global min-max across band normalization using the 1st and 99th percentile image values. See
here for code and
configuration.
Performance metrics on the validation split (calculated across samples drawn from countries listed below):
