Limited Error Raster Compression Support

[micahcc]

For certain cases lossy compression with limited error makes sense, e.g. heightmap data. LERC is a standard way of handling this (https://github.com/esri/lerc) in geotiffs (https://gdal.org/en/stable/drivers/raster/gtiff.html#overviews).

Made PR to implement here:
#409

[micahcc]

Actually after further testing LERC seems worse compression for my heightmaps than scaling + casting to int16 and then adding a ScaleZ to account for the scaling, so I probably don't need this.

while YMMV it doesn't seem like LERC helps in my case

while YMMV it doesn't seem like LERC helps in my case

DEM Compression Benchmarks

Source: Copernicus_DSM_10_N00_00_E035_00_DEM.tif (3601x3601, Float32, uncompressed = 50MB)

Single Source Tile Compression

• Int16 + LERC_ZSTD lossless: round to int16, losing 0.5m precision, then apply
  LERC_ZSTD lossless
• Int16 + DEFLATE: round to int16, losing 0.5m precision, then apply DEFLATE
  compression
• Int16/X: multiply the floating values by X, cast to int16, then add ScaleZ of
  1/X so that the values may be interpeted as their original value

Approach	Size	Compression	Precision
Float32 raw (uncompressed)	50M	1.0x	lossless
Float32 + LERC lossless	44M	1.1x	lossless
Float32 + DEFLATE + predictor	35M	1.4x	lossless
Float32 + ZSTD + predictor	35M	1.4x	lossless
Float32 + ZSTD + float predictor	29M	1.7x	lossless
Float32 + LERC_ZSTD 10cm	19M	2.6x	$\pm 10cm$
Float32 + LERC_ZSTD 50cm	15M	3.3x	$\pm 50cm$
Float32 + LERC_ZSTD 1m	14M	3.6x	$\pm 100cm$
Int16 + LERC_ZSTD lossless	12M	4.2x	$\pm 500cm$
Float32 + LERC_ZSTD 5m	10M	5.0x	$\pm 500cm$
Float32 + LERC_ZSTD 10m	9M	5.6x	$\pm 1000m$
Int16 + ZSTD + predictor	8.1M	6.2x	$\pm 50m$
Int16 + DEFLATE + predictor	7.9M	6.3x	$\pm 50m$
Int16/2 + DEFLATE + predictor	6.1M	8.2x	$\pm 100cm$
Int16/4 + DEFLATE + predictor	4.5M	11.1x	$\pm 200cm$