Recently, I’ve been working on developing self-supervised deep learning models for raster processing (specifically processing DEMs). Generating labelled training data as part of the self-supervised learning process involves the manipulation of 10s to 100s of thousands of rasters. In the Python ecosystem (which is currently the dominant language for developing deep learning methodologies) currently available raster processing libraries (such as GDAL) are prohibitively slow for performing such processing volumes.
To overcome this, I developed a new Python library for raster manipulation and I/O, named FasterRaster. The goal of FasterRaster was to implement fast parallel raster processing operations via Cython wrapped OpenMP C functions. Memory allocation, I/O, and data objects are handled in NumPy allowing for straightforward memory management and easy addition of processing operations.
You can check out the github repository here.