I was thinking of using Google Earth Engine to run a relatively complex ML model (tensorflow based), on high resolution tiles (RGB). About 0.5m/px input would be the best. I need to process entire Europe landmass and some captured seas.
The model does image recognition / classification and localization. The input are about 250x250 px aerial or satellite images (RGB). I want to process the entire region of interest in overlapping tiles, so I don't miss any classified object in the region of interest. I have 5 about labels. All objects are spatially small, but still has spatial dimension, so the processing / classification can't be done at lower spatial resolution, or on per-pixel or smaller blocks basis.
My rough estimate, is for Europe it is about 2.8 billion tiles (700 million tiles without overlap) to classify. This looks doable but obviously an enormous processing task. The model is very fast and takes maybe 10 milliseconds to run on a CPU (possible less). That is about 7800 CPU-hours, which with few big servers can be done in few days.
The result of the program would be list of about 20000 tiles with labels. (Rest would be label=none, to be ignored).
Is this something that can be done with Google Earth Engine actually? If not, is it possible to access Google Earth Engine or Google Maps tiles directly from Google Cloud Compute maybe with reasonable licensing conditions? Maybe Bing?