geotrellis/geotrellis

GeoTrellis is a geographic data processing engine for high performance applications.

Simeon H.K. Fitch

@metasim

WRT @vpipkt 's question/analysis, why is this block of code necessary?
https://github.com/s22s/geotrellis/blob/9d8ceffc435906a1d18e1283bda5e22b1b4f37e4/gdal/src/main/scala/geotrellis/raster/gdal/GDALUtils.scala#L45-L57
Does GDAL not have a built-in indicator for unsigned byte cell type?

The issue at hand is that we want all the metadata about scene/tile, including the cell type, and only want to be reading the header. This mim/max scan of the cells triggers a larger read of the data, eliminating the ability to reason about (e.g. filter) a tile on metadata alone.

Grigory

@pomadchin

@metasim yes, to determine the celltype correctly we need bitsPerSample and sampleFormat, bitsPerSample is number of bits per tiff sample, and the sampleFormat is used to determin the ‘sing’ of values.

LIBTIFF exposes this information and in geotiff native jvm reader we also use this information to determine the cellType properly: https://github.com/locationtech/geotrellis/blob/7871e015129ae4a1ccfd06a23544a6de9ac68c6e/raster/src/main/scala/geotrellis/raster/io/geotiff/BandTypes.scala#L37

However GDAL doesn’t expose such information, or I didn’t find a proper way of doing it. Probably you can find smth in the GDAL C API and somehow get the information from the used driver about the raster cellType, but I’m not sure that it is possible to expose such information via GDAL

Simeon H.K. Fitch

@metasim

Ouch... what a PITA!

Grigory

@pomadchin

yea, without this ‘conversion function’ GDALRasterSource and all GeoTrellis types would behave really differently

Simeon H.K. Fitch

@metasim

I knew it was in GeoTIFF based on the JVM reader.

Surprised GDAL didn't handle it more gracefully.

Grigory

@pomadchin

they just use different types for tiffs

Simeon H.K. Fitch

@metasim

I'd have expected the NoData handling part... makes sense

I think @vpipkt has a workaround that will keep min/max from being computed for all other cell types (because the parameter is call-by-name), but still still happen for byte types.

Grigory

@pomadchin

¯\_(ツ)_/¯ I would recommend also to doubleck the GDAL API - mb there is smth we can pull this information from

Simeon H.K. Fitch

@metasim

will do

After I finish the GT 3.x upgrade ;-)

Grigory

@pomadchin

: D

yo @kervel to make this function work you need only to have import geotrellis.spark._ in the call scope. If you have it and function still doesnt work try to use withGetBoundsMethod(rdd).getGridBounds in your code, this at least will give you a readble compile time error.

Jason T Brown

@vpipkt

My reading of this GDAL doc is that there will never be a signed Byte type. And that would mean that the first case here is always true and the second is never evaluated?

Grigory

@pomadchin

@vpipkt to follow the native JVM reader both cases can be computed https://github.com/locationtech/geotrellis/blob/7871e015129ae4a1ccfd06a23544a6de9ac68c6e/raster/src/main/scala/geotrellis/raster/io/geotiff/reader/GeoTiffInfo.scala#L53

we have a bit more rich typesystem; it is wider than what GDAL Provides

also we just store data differently; we don’t have unsugned types in java (: so any byte tiff is by default in [-128; 127] ranges; and UByte allows to interpret them as [0..255] values

Grigory

@pomadchin

ha @vpipkt mb you’re right

Frank Dekervel

@kervel

@pomadchin I know where the error is coming from.. the withGetBoundsMethod Class takes the rdd as instance variable, and needs to be serializable because of this

Grigory

@pomadchin

Ah….. nd value probably can be below zero… because we don’t have ubytes

but still check it i.e. what if NoData Value=255

Frank Dekervel

@kervel

Just like the other implicit classes in the same file, which already are.

Grigory

@pomadchin

@kervel mb; do you have serialziation issues?

Frank Dekervel

@kervel

Yes. But the fix is simple

Grigory

@pomadchin

Nice! would you like to create a PR with marking this implicit class as serializable?

Frank Dekervel

@kervel

Ok.

Grigory

@pomadchin

:100:

It would be awesome if you could sign an ECA https://github.com/locationtech/geotrellis/blob/master/docs/CONTRIBUTING.rst#eclipse-contributor-agreement-eca

after that commit changes via git commit -s -m “commit message” command

Frank Dekervel

@kervel

Ok.

Jason T Brown

@vpipkt

@pomadchin i am not sure how to proceed, but my understanding is that since GDAL's type system is narrower, a GDT_Byte should always result in either UByteCellType, UByteUserDefinedNoDataCellType or UByteConstantNoDataCellType

the function as called will never return a BitCellType either because no typeSizeInBits is passed

(at least as called in this context)

Jason T Brown

@vpipkt

@pomadchin happy to keep the discussion going on this PR: locationtech/geotrellis#3150

Grigory

@pomadchin

@vpipkt can you cover it with test? so we would know that GDALRasterSource behaves the same way GeoTiffRasterSource does

Simeon H.K. Fitch

@metasim

If I have, say, an RDD[ProjectedRaster[Int]] (heterogeneous in CRS) and want to write them all out to HDFS, isn't there an "unstructured cog writer" feature in GT?

Or rather, is that the right/best option?

I don't want to reproject anything or resample it into a new grid.

Grigory

@pomadchin

@metasim nope we didn’t solve it :s

There it still a problem how would you collect metadata of multiple raster sources

Simeon H.K. Fitch

@metasim

Did you do some experimental work on it?... seem to remember some code somewhere ages ago....

Simeon H.K. Fitch

@metasim

where you were writing out debugging vrt files? or am I making that up.

Grigory

@pomadchin

@metasim we didnt resolve that; imagine the case that you have multiple RasterSources
you want to figure out how to tile them to some layout in some projection…

it would mean that you probably need to collect metadata in some CRS or reprojet it into some CRS to have some idea of the entire rasters list extent

ahhhh youre talking about persisting

and not about reading

hmmm can you describe your usecase a lil bit more?

Im jsut flygin in a different context at the moment

Simeon H.K. Fitch

@metasim

yeh, sorry. Say you want to write out a bunch of DL chips as geotiffs, organized via some arbitrary partitioning scheme. Maybe some json sidecars for metadata (or not).

BTW, progress is being made: https://circleci.com/gh/s22s/rasterframes/1757

Where communities thrive

People

Repo info

Activity