this is fine for now since we don't use it for anything in the ML pipeline for this project or in the past, but we should consider investing time in georeferencing these. It wold let us know where in the world we have annotations, and let us inspect model performance geographically.
Adding this info back would entail calculating the projection info for each image in the test, val, and training sets. The imagery is resampled, so the projection info that is saved needs to take that into account. This could be saved in place. Or, replacing these geotiffs with georeferenced tiffs from https://registry.opendata.aws/sentinel-1/ which is the original source and resampling these the same way so we don't lose the correspondence with the annotations.
this is fine for now since we don't use it for anything in the ML pipeline for this project or in the past, but we should consider investing time in georeferencing these. It wold let us know where in the world we have annotations, and let us inspect model performance geographically.
Adding this info back would entail calculating the projection info for each image in the test, val, and training sets. The imagery is resampled, so the projection info that is saved needs to take that into account. This could be saved in place. Or, replacing these geotiffs with georeferenced tiffs from https://registry.opendata.aws/sentinel-1/ which is the original source and resampling these the same way so we don't lose the correspondence with the annotations.
CC @srmsoumya