Enhancing RGB Tile Contrast

[ChristopherKotthoff]

Summary

This pull request enhances the image preprocessing capabilities in detectree2/preprocessing/tiling.py by introducing an optional RGB contrast enhancement feature and fixing some inconsistencies with return values across the tiling functions.

Changes

RGB Contrast Enhancement

I've added an enhance_rgb_contrast parameter to the process_tile, process_tile_train, and tile_data functions. When this feature is enabled, the system rescales RGB pixel values from their original distribution to a 1-255 range, reserving 0 for nodata values. The contrast enhancement uses percentile-based rescaling (0.2% to 99.8%) to increase RGB diversity, which significantly improves the accuracy of jungle crown delineations.

Context Manager Fix

There was also a resource handling issue I discovered where functions were returning a data object from within a context manager (with rasterio.open(img_path) as data:), which could lead to problems down the line. I've fixed this by having the functions return only the out_transform from the data object, since that's all the calling methods actually need anyway. This change affects process_tile, process_tile_ms, and process_tile_train functions.

Minor Adjustments

Additionally, I made a small adjustment to the convex mask precision by reducing the buffer size from 5 to 3 for the convex delineation mask. This value is still large enough to be effective while avoiding too much inclusion of nearby tree crowns.

Future Considerations

I'll admit the function signatures are getting pretty parameter-heavy with these additions. I'm planning a follow-up pull request in the near future to refactor and clean up these function interfaces, but the current implementation is still manageable for now.

Compatibility

Existing workflows will work exactly as before. The parameter enhance_rgb_contrast defaults to true, allowing for immediate performance gains if the data is re-tiled.

[ChristopherKotthoff]

Will be marked as ready for review after more testing.

[ChristopherKotthoff]

Many problems are arising with dependency issues and module version mismatch in the GitHub actions. MyPy was raising errors that I was not able to replicate on my system with the same version numbers. I was able to fix them via commit fix -> overvation -> commit fix -> etc.
Also it seems that the Docker Action never is dispatched.

But for this pull request: All tested, everything works as it should.
Please feel free to review / merge.

[PatBall1]

Looks good to me