Splitting rasters into image chips for deep learning

[williamlidberg]

Something I find myself doing a lot is preparing raster data for convolutional neural networks. My current method is to use splittraster: https://github.com/williamlidberg/Cultural-remains/blob/main/tools/split_training_data.py The great part of this method is that it continues to name files based on previous files in the output directory. The downside is that the output files do not have a coordinate system/projection.

ArcGIS Pro has a tool to export data as image chips with coordinates/projection but ArcGIS Pro does not work in our docker containers and also requires a rather expensive license.

My suggestion would be a tool that split input tiles into smaller image chips that can be used for deep learning. optimally the tool would come up with logical names of the output files so they can be looped over multiple input files without overwriting the outputs from previous loops. It would also be awesome if the split images retained their coordinates so they can be inspected or worked on later.