This is spot on! A key part of the challenge is to develop more robust models working across a range of resolutions, imagery capture conditions, diverse geographies. Your ideas to incorporate training data at native resolutions and of different sizes (256x256 up to 1024x1024) could be promising, and could be achieved using rasterio’s windowed reads: Windowed reading and writing — rasterio 1.4dev documentation
In general, windowed reads should offer some more flexibility to chip the image rasters to whatever size and shape you like. There’s a small example of a 1024x1024 window read at native resolution at the bottom of the pystac starter colab notebook provided in the STAC Resources competition page.
One other thing to be aware of if you’re creating training chips to standardized webmap tile zoom levels and squares is that there’s almost always some resampling and reprojection (affine transform) happening from the original image as a resut. Not saying if that’s good or bad for your model performance…I don’t know, worth experimenting! For instance, here’s how rio-tiler works under the hood: https://github.com/cogeotiff/rio-tiler/blob/6b3b57e4d837cfa855d9b4a74175812339ac433e/rio_tiler/utils.py#L337