Problem
GDAL writes tiled JPEG TIFFs (compress=JPEG photometric=YCBCR) where each tile is a JPEG fragment that depends on shared quantization and Huffman tables stored once in the file under tag 347 (JPEGTables). The reader does not inject those tables into each per-tile stream, so reading any GDAL-written tiled JPEG TIFF raises:
OSError: broken data stream when reading image file
xrspatial-written JPEG TIFFs happen to be self-contained per tile, which is why the existing test suite does not catch this.
Where
xrspatial/geotiff/_compression.py:820-829 — JPEG decode path.
Repro
import rasterio, numpy as np
from xrspatial.geotiff import open_geotiff
arr = (np.random.rand(512, 512, 3) * 255).astype('uint8')
with rasterio.open('/tmp/jpegtiled.tif', 'w', driver='GTiff',
height=512, width=512, count=3, dtype='uint8',
tiled=True, blockxsize=256, blockysize=256,
compress='JPEG', photometric='YCBCR') as dst:
dst.write(np.transpose(arr, (2, 0, 1)))
open_geotiff('/tmp/jpegtiled.tif') # OSErrorFix sketch
Read tag 347 (JPEGTables) once at IFD parse time. Before passing each tile's JPEG bytes to the decoder, prepend the tables (or splice them in at the right SOI offset, since JPEGTables ends with EOI and each tile starts with SOI — drop the tile's SOI and the tables' EOI and concatenate).
Severity
Interop break, not numerical drift. Discovered during geotiff accuracy sweep but deferred from that sweep because pixel-value semantics are correct once the JPEG decodes.
Problem
GDAL writes tiled JPEG TIFFs (
compress=JPEG photometric=YCBCR) where each tile is a JPEG fragment that depends on shared quantization and Huffman tables stored once in the file under tag 347 (JPEGTables). The reader does not inject those tables into each per-tile stream, so reading any GDAL-written tiled JPEG TIFF raises:xrspatial-written JPEG TIFFs happen to be self-contained per tile, which is why the existing test suite does not catch this.
Where
xrspatial/geotiff/_compression.py:820-829— JPEG decode path.Repro
Fix sketch
Read tag 347 (
JPEGTables) once at IFD parse time. Before passing each tile's JPEG bytes to the decoder, prepend the tables (or splice them in at the right SOI offset, sinceJPEGTablesends with EOI and each tile starts with SOI — drop the tile's SOI and the tables' EOI and concatenate).Severity
Interop break, not numerical drift. Discovered during geotiff accuracy sweep but deferred from that sweep because pixel-value semantics are correct once the JPEG decodes.