SAR_CEOS driver used instead of CEOS driver for certain datasets

[rprinceley]

The SAR_CEOS driver incorrectly identifies and attempts to open certain CEOS datasets. A workaround would be to change the driver order, but ideally, SAR_CEOS should have a proper Identify implementation (or more checks).

​CEOS2001.I3B

[antonio]

There is something strange.

In [1]: import struct

In [2]: s = struct.Struct('>I4BI')

In [3]: fd = open('CEOS2001.I3B', 'rb')

In [4]: h = fd.read(s.size)

In [5]: s.unpack(h)
Out[5]: (1, 63, 192, 18, 18, 4136)

In [6]: fd.seek(4136)

In [7]: h2 = fd.read(s.size)

In [8]: s.unpack(h2)
Out[8]: (538976288, 32, 32, 32, 32, 538976288)

Can you provide more info about the test dataset?

[antonio]

Just to be sure

$ md5sum CEOS2001.I3B 
f4e0533fe653cf76eff53e339015726e  CEOS2001.I3B

[rprinceley]

Antonio,

This raster is part of a sample CD from ​Earth Remote Sensing Data Analysis Center (ERSDAC). I have examined the headers of other rasters in this dataset, they seem to be fine. We can conclude that this is a bad file as you suspect.

The fact that CEOS driver reads this file without any error or warning (even though pixels are read incorrectly) is a problem. GDAL should generate a error in this case.

[warmerdam]

I am digging into this. Note that the file does seem to more-or-less work with the generic CEOS driver despite what appears to be some corruption problems. I will see what I can do to have the SAR_CEOS driver skip it.

[warmerdam]

The problem with the file is that the record length encoded in the ceos record header does not reflect the actual offset between records in the file. This likely is due to the way that the file was created from variable length tape.

I have adjusted the SAR_CEOS driver to check the record sequence numbers (a good indicator if we are getting what we expect). If the sequence number is wrong for the second record of the imagery file we just back out of the driver quietly (with a debug message). In other places we error out.

This gives us the desired fallthrough to the regular CEOS driver. Note that the record ceos driver doesn't really work properly either - the imagery is partly corrupted because the offset to the start of imagery is a bit wrong but it is still somewhat useful.

I have applied the change in trunk (r22662) and in 1.8 branch (r22663).

Robin - I'll let you back port it to 1.8-esri or whereever else you might want it. Thanks for the report.