Training and inference pipeline

Our goal is to create a well-documented and flexible pipeline. Ideally, we should be able to specify some configuration parameters as inputs and automatically generate the desired results.

A pipeline to create d-pixels should be developed (independent of the Barlow Twins main pipeline) that takes care of geospatial alignment issues, the use of cloud masking, and the ability to use both raster and vector datasets. It will need to load and handle large time series datasets. This pipeline should produce time series either pixel-wise or patch-wise features to allow for representation outputs to be placed into a spatially-oriented raster (i.e to create a wall-to-wall representation map). Additionally, the pixel- or patch-wise features should be in a format usable for other ML approaches.