Inference using irregularly sampled time series
When a model is trained with the kind of data prepared by the sampling described here, the inference step has to be adapted as follows.
- Concatenate the images as for the sampling step (no gapfilling nor feature extraction)
- Concatenate the masks as for the sampling step
- Provide these stacks and the corresponding date files (like the ones used for the gapfilling) to the ML model
- Provide also the list of all available dates (for all tiles)
The model is responsible for the formatting of the data so that it corresponds to what was used for the training.
One point to consider is that it may be useful to concatenate the images and the masks so that we can use what has been done for the inference with scikit-learn.