Trouble Hiding NoData Value

[dseminara-pivot]

Hello there!

I just started experimenting with titiler yesterday and have made great progress in a rather short amount of time (deploying it to the cloud, rendering tiles in a variety of web mapping libraries, etc)! However, I can't seem to generate tiles that respect our nodata value of -10000 and I'm hoping this is something as simple as me still being new to titiler.

In QGIS, it seems to automatically 'know' this value and drops it from view. Similarly, when using the same COG image outside of Titiler and directly from the source - I'm able to tell OpenLayers what the value is and have it appropriately remove it from the scene.

I've also tried adding &nodata=-10000 to the end of the url and have used the swagger docs to generate it for me as well - but it doesn't seem to respect that param.

Attached is a gdalinfo output (minus location info like Origin and Corner Coordinates) and a small snippet of what I'm referring to. I saw that rio faux could be provided as an option, but unfortunately I can't provide any spatial location to where this image is from.

> gdalinfo out.tif
Driver: GTiff/GeoTIFF
Files: out.tif
Size is 11557, 10545
Coordinate System is:
GEOGCS["WGS 84",
    DATUM["WGS_1984",
        SPHEROID["WGS 84",6378137,298.257223563,
            AUTHORITY["EPSG","7030"]],
        AUTHORITY["EPSG","6326"]],
    PRIMEM["Greenwich",0],
    UNIT["degree",0.0174532925199433],
    AUTHORITY["EPSG","4326"]]
Pixel Size = (0.000000141738231,-0.000000141738231)
Metadata:
  AREA_OR_POINT=Area
  OVR_RESAMPLING_ALG=NEAREST
Image Structure Metadata:
  COMPRESSION=DEFLATE
  INTERLEAVE=PIXEL
Band 1 Block=512x512 Type=Float32, ColorInterp=Blue
  Description = Band1
  NoData Value=-10000
  Overviews: 5779x5273, 2890x2637, 1445x1319, 723x660, 362x330
Band 2 Block=512x512 Type=Float32, ColorInterp=Green
  Description = Band2
  NoData Value=-10000
  Overviews: 5779x5273, 2890x2637, 1445x1319, 723x660, 362x330
Band 3 Block=512x512 Type=Float32, ColorInterp=Red
  Description = Band3
  NoData Value=-10000
  Overviews: 5779x5273, 2890x2637, 1445x1319, 723x660, 362x330
Band 4 Block=512x512 Type=Float32, ColorInterp=Undefined
  Description = Band4
  NoData Value=-10000
  Overviews: 5779x5273, 2890x2637, 1445x1319, 723x660, 362x330
Band 5 Block=512x512 Type=Float32, ColorInterp=Undefined
  Description = Band5
  NoData Value=-10000
  Overviews: 5779x5273, 2890x2637, 1445x1319, 723x660, 362x330
Band 6 Block=512x512 Type=Float32, ColorInterp=Undefined
  Description = Band6
  NoData Value=-10000
  Overviews: 5779x5273, 2890x2637, 1445x1319, 723x660, 362x330
Band 7 Block=512x512 Type=Float32, ColorInterp=Undefined
  Description = Band7
  NoData Value=-10000
  Overviews: 5779x5273, 2890x2637, 1445x1319, 723x660, 362x330

Tile Request in the browser:

Using /cog/preview

[vincentsarago]

👋 @dseminara-pivot

Without access to the file it will be pretty hard for me to help.

If there is an issue, it will be in the rio-tiler side anyway. TO debug this you could just try to play with rio-tiler

from rio_tiler.io import Reader

with Reader("my.tif") as src:
    im = src.preview(indexes=1)  # Read the first band
    
    data = im.data
    mask = im.mask  #<-- this is the part you want to check to make sure the nodata is used to create the mask

Note: we are currently working on the next version of rio-tiler which change a lot of things for the mask creation (switch from a per-dataset to a per-band mask) and also fixing some issue when working with GCPS referenced files.

[dseminara-pivot]

Thanks for the quick reply @vincentsarago! ..and totally understand the difficulties with assistance without access to our image dataset. Does perhaps the rescale param round? Our pixel values are incredibly low with rescale=0,0.006. I've tried modifying the min value to be greater than 0 but below our max of 0.006 but get the same result.

Looking at the file within Reader, i see data with the matching -10000.0 value for what should be expect to be nodata and within mask a value of 0

[vincentsarago]

the mask value should be either 0 or 255 (0 mean masked, 255 means no masked)

[dseminara-pivot]

Perhaps asking this another way - in hopes I'm not misunderstanding what the mask does. Does/can the mask provide a transparent 'value' similar to when a tile is at edge of the actual image. Basically, I'm hoping that the black area which represents nodata can transparent as well.

[vincentsarago]

Yes it should be transparent as well.

if you cannot share the data publicly but can directly to me, please feel free to contact me at [email protected]

[vincentsarago]

after inspecting the data I think I know what's going on!

The band 6 and 7 do not have the same nodata mask. Their value is set to 0 while the other bands are at -10000

Sadly the current version of rio-tiler use a per-dataset mask, meaning that it will construct a mask using the sum of all the mask, and because the band 6 and 7 says it shouldn't mask the area on the right of the image, the combined mask return 255 instead of 0.

https://github.com/cogeotiff/rio-tiler/blob/main/rio_tiler/reader.py#L197-L202

Next version of rio-tiler will use per-band mask so it should better handle this case (ref: cogeotiff/rio-tiler#579)

https://github.com/cogeotiff/rio-tiler/blob/30dc83cdc3154ac2820363d33f1ef6ced8cb8d25/rio_tiler/reader.py#L222-L233

Possible fix

On your side, while we wait for new rio-tiler/titiler, you could edit your file to have a kinda unified nodata area (editing band 6 and 7 to have the same values as band 1 where it's -10000)

[dseminara-pivot]

Thanks Vincent. That makes sense...our last two bands are based on derived values which probably explains the mismatch. Appreciate you investigating.