Make data augmentation available for the regression
Objective
Currently, the data augmentation is only available for the classification. It can be useful to make it available for the regression too.
Context
The current augmentation steps is based on the OTB application. It requires the field name in the input vector file providing the class membership. For regression, we do not have this kind of field (we have continuous values to be inferred).
Proposed solution
We can create an additional field in the input vector file that assign every sample into a group/class that belongs to an interval of values, similar to what it is done when constructing an histogram, and then give this field to the otb app.
For regression, a fake column of unique value is already created to split polygons into train/test subset. We can define here a function creating fake classes, using for instance numpy.digitize. Then using sample augmentation step should be straightforward.
Several strategies can be defined for the creation of fake classes. From our experience in PARCELLE, equal size intervals work quite well. So if we want 10 fake classes, the range of values to be regressed can be divided into 10 continuous intervals of same length.
An option could be defined to use as intervals as unique values to be regressed for extreme cases.