Generation of sample files for irregular time series
In order to feed ML models which are able to work with irregularly sampled time series, we need to add a new workflow for the sampling step.
- Use the SampleSelection step as usual to generate the positions of the sampled pixels
- Concatenate all the images of a tile as usual, but do not use gapfilling nor feature extraction
- Use the SampleExtraction on the concatenated images and the masks in order to generate a sample file where the features are the image and mask values for each date
- Since each tile may have different acquisition dates, the fusion of the sample files for all tiles must take this into account: the features of the fused file will be the union of all the features of all tiles. This needs to put a flag value for the columns for which a particular pixel has no real acquisition date. I propose to set reflectances to 0 and masks to a 3rd value (valid, masked, unavailable).
This workflow delegates all feature computation (NDVI, etc) to the ML model.