Closed Jack-bo1220 closed 1 year ago
They are independent in our training. We use public domain NAIP imagery for high-resolution and those are only available in the continental US, but for low-resolution we sample tiles from around the world.
The vast majority of tiles with high-resolution imagery are also covered by low-resolution imagery, so it should be feasible to train on only tiles with both images (Sentinel-2 and NAIP) available. However we find that Sentinel-2 does not provide any extra information (at least for the 137 Satlas labels) that NAIP doesn't already provide, since NAIP is 10x+ higher resolution.
Are low resolution images and high-resolution images one-to-one corresponding? Or are they independent?