Edges of dataset load black patches

As all the tiles are warped to one crs I believe the edge of the dataset loads black patches. I attached an example of a warped NAIP tile that is rotated because of the warping. The BoundingBox is based on min & max values therefore capures more than the actual outline of the tile.

Is there a good way to handle this? Maybe shrinking the BoundingBox somehow to only contain data.

PS: This is also an example of why loading just the hit in #547 does not work even for datasets perpendicular to their crs. Will have to investigate how the merge handles this.

[adamjstewart]

It depends on if your goal is training or prediction. For training, yes, it’s fine to lose a bit of data assuming you have a lot of data. For prediction, you want to make a prediction for every pixel on the map, not just the ones near the center of each image tile. I agree that this would be a useful addition though. I’m not sure if there’s a way to automatically predict the maximum size of the bounding box within the image that contains data. I guess it depends on the metadata available for that particular satellite.

[Modexus]

Ah yes predicting, I didn’t think of that as I only really train for research.

Are the landsat images already rotated or is it the same effect from warping? It may be feasible to warp the coordinates of the bounds and use the smaller value for the bounding box.

If they are already rotated is it possible to warp the bounds back to the original crs (if known) and then warp the coordinates to the new crs?