[R-sig-Geo] writing shapefiles / DBF files when input data contains NA

[Roger Bivand]

On Wed, 8 Oct 2008, Dylan Beaudette wrote:

> On Tuesday 07 October 2008, Dylan Beaudette wrote:
>> On Tuesday 07 October 2008, Roger Bivand wrote:
>>> On Mon, 6 Oct 2008, Dylan Beaudette wrote:
>>>> Hi,
>>>>
>>>> I have noticed that saving data to files that include a DBF, result in
>>>> bogus data where there were NA. Using the write.dbf() function from
>>>> the foreign package seems to work a little better, but I still get odd
>>>> results in numeric columns. Writing to GRASS with the methods in the
>>>> spgrass6 package results in some thing that looks like this:
>>>
>>> Dylan,
>>>
>>> I'm afraid that there is no good solution for this at all. DBF does not
>>> seem to have a clear and uniform NA treatment (or even !is.finite()
>>> treatment). The only work-around is to preprocess the data.frame in the
>>> output object to insert known NODATA values, and to replace those flags
>>> manually on the GRASS side. This could possibly be written as a wrapper
>>> around writeVECT6(). The help page does say:
>>>
>>>      "Please note that the OGR drivers used may not handle missing data
>>>       gracefully, and be prepared to have to correct for this manually.
>>>       For example use of the 'readOGR' PostGIS driver directly may
>>>       perform better than moving the data through the DBF driver used in
>>>       this function - or a PostgreSQL driver used directly or through
>>>       ODBC may be a solution. Do not rely on missing values of vector
>>>       data moving smoothly across the interface."
>>>
>>> I did try to look at the SQLite driver on the GRASS side, which might be
>>> more robust, but did not see how to proceed.
>>>
>>> One possibility is not to recode, but to build an NA mask on the R side,
>>> and then loop over fields on the GRASS side for the chosen driver
>>> inserting NAs in the correct rows (whatever the syntax for that might
>>> be). Would this be db.execute with an insertion of SQL NULL?
>>>
>>> Can we redirect this discussion to the statgrass list, because GRASS
>>> developers follow that list?
>>>
>>> Best wishes,
>>>
>>> Roger
>>
>> Sorry for the cross-posting. Wanted to clarify where this thread is
>> going/went.
>>
>> Hi Roger--
>>
>> It looks like the limiting factor in this equation is the code used in
>> v.out.ogr.
>>
>>> From the GRASS-dev + Frank W's help:
>>>> Sounds good :)
>>>> Does anyone know how to fix
>>>>  vector/v.out.ogr/main.c
>>>> to support NULLs? I see db_set_value_null() in
>>>>  lib/db/dbmi_base/value.c
>>>> which might be relevant.
>>>
>>> Markus,
>>>
>>> Once you establish which GRASS attributes are NULL, you can ensure they
>>> are pushed out to OGR as null by just skipping the step that sets them.
>>> Perhaps that will help a bit.
>>
>> So, once v.out.ogr is fixed, this should clear up several issues:
>>
>> 1. import of vector data into R via spgrass6 methods
>> 2. better compatibility of vector data exported from GRASS
>>
>> I still do not know why writeOGR() does not create correct DBF files... it
>> may be related to the code in v.out.ogr....
>>
>> Cheers,
>>
>> Dylan
>
> Some follow-up: the incorrect handling of NULL values appears to be related to
> the current implementation of v.out.ogr AND readOGR() / writeOGR().

OK, this makes sense, because parts of readOGR() / writeOGR() were written 
based on the logic of v.in.ogr and v.out.ogr, and more attention was given 
to the geometries than the attribute fields. If the GRASS code was taking 
liberties with handling NAs, then that behaviour is very probably present 
in readOGR() / writeOGR() too.

The rgdal package has a public sourceforge CVS repository, so everybody 
please feel free to browse for bugs. It would be helpful to have a set of 
vector files with valid NAs (not just shapefiles), and a set of sp objects 
with NAs, and to be able to move them in and out of both R and GRASS (and 
other software) with the NAs intact.

As a first bite, OGRFeature::IsFieldSet() seems to test whether the field 
is set or not. It isn't used in ogrReadColumn() in src/ogrsource.cpp in 
rgdal, nor the equivalent in OGR_write() in src/OGR_write.cpp.

Assuming that we can correct these to use OGR NULL data representations 
(would that be unset the field for the feature?), we then depend on the 
drivers using the same logic. In addition, non-OGR written files need to 
use the same understanding of NULL as the OGR drivers. GRASS v.in.ogr() 
does use OGR_F_IsFieldSet(), and if not set writes a NULL to numeric 
fields and an empty string to the others. Fixing writeOGR() ought to get 
NAs from R to GRASS. v.out.ogr does not seem to use OGR_F_UnsetField() on 
the fields being output, and readOGR() does not test for the fields being 
unset either - so getting NAs from GRASS to R needs more work.

This is described in extenso here because things don't happen by 
themselves, and this particular overlap of R/OGR/GRASS code probably 
matters to regular users of rgdal. Collaboration in fixing the handling of 
NAs in vector data files invited!

Roger

>
> Dylan
>
>

-- 
Roger Bivand
Economic Geography Section, Department of Economics, Norwegian School of
Economics and Business Administration, Helleveien 30, N-5045 Bergen,
Norway. voice: +47 55 95 93 55; fax +47 55 95 95 43
e-mail: Roger.Bivand at nhh.no