balance the training set (Maaike)

[sonjageorgievska]

the training set should be balanced to have roughly equal number of cutouts that have a label = 0 (no damage) and label >0 (some damage). This can be done by removing the central parts of Antartica where there are no iceshelfs and/or by using metadata that indicates damaged areas.

[meiertgrootes]

while this has been done for the small exploratory data set, other means must be invesigated for the full set.

is ice shelf selection sufficient if not, select based on independent models predictions Or use model input as basis for creating balanced data sets