Georeferenced predictions

[lcoandrade]

Is there any advances towards a transparent way to make georeferenced predictions? So far, I'm working with Rasterio to accomplish that. The problem is that my process is taking way longer than making the georeferenced prediction using RasterVision, for instance.

[adamjstewart]

Not yet, but this is something I would really like to see if anyone wants to try to take a stab at it. #1407 and #1440 contain a list of hurdles and suggestions for this. If we can keep the CRS/transform and other metadata in the R-tree and pass it through Lightning, we should be able to implement a predict_step that saves the output to a GeoTIFF.

[lcoandrade]

I haven't taken a look at how RasterVision does that. Do you know what was their approach?

[adamjstewart]

Not sure how they do it, maybe @lewfish @AdeelH can comment

[AdeelH]

For semantic segmentation, Raster Vision aggregates the predictions for each window in a numpy array and then writes that array out to a GeoTIFF with the same extent, projection etc. as the input GeoTiff. This functionality is spread across SemanticSegmentationLabels and SemanticSegmentationLabelStore.

It should theoretically be possible to use TorchGeo-produced predictions with this functionality by appropriately modifying this example. You would need an iterable of Raster Vision windows, an iterable of corresponding predictions, a ClassConfig, and a CRSTransformer.

P.S. @adamjstewart, FYI Lewis is no longer involved in the project.

[lcoandrade]

The idea I'm using so far is this (using Rasterio from_origin and merge):

1. I create a Grid Dataloader (RasterDataset+GridGeoSampler+Dataloader) (something parallel to SlidingWindowGeoDataset);
2. For each sample, I create a geotiff with the CRS (from the dataset) and geotransform matrix using from_origin(bbox.minx, bbox.maxy, pixel_size, pixel_size);
3. I add each chipped geotiff to a list;
4. I merge the list of geotiffs to make the full geotiff prediction;
5. Save this as a file.

Do you see a bottleneck in this approach?