[GRASS-user] Precision of raster/vector float values

[Markus Metz]

Another idea, maybe it helps in your case:

You can use raster values as z values in vector points (r.to.vect -z)
which would be in case of a DCELL raster the same like the raster
values because both are double precision floating point.

Markus M


On Thu, Aug 23, 2012 at 4:10 PM, Johannes Radinger
<johannesradinger at gmail.com> wrote:
> On Thu, Aug 23, 2012 at 3:23 PM, Markus Metz
> <markus.metz.giswork at gmail.com> wrote:
>> On Thu, Aug 23, 2012 at 2:40 PM, Johannes Radinger
>> <johannesradinger at gmail.com> wrote:
>>>>> I have a GRASS raster map which contains small values (e.g 2.5e-15) among
>>>>> others.
>>>>
>>>> This is (dangerously) close to the IEEE limit of double precision
>>>> floating point values. If GRASS tests for fp precision, it regards
>>>> (absolute) values smaller than 1.0e-15 (GRASS_EPSILON) as zero.
>>>>
>>>> v.what.rast prints fp values with 10 decimal places. The equivalent
>>>> issue has aready been fixed in d.what.rast, now also for v.what.rast
>>>> in trunk r52857. For FCELL maps, it prints 7 decimal places, for DCELL
>>>> maps it prints 15 decimal places. Printing means here that fp values
>>>> are first converted to text, then uploaded to the attribute table.
>>>>
>>>
>>> So if I understand you correctly: 1.0e-15 is the smallest number for
>>> values in attribute columns although the values in a DCELL raster
>>> can be  smaller? A bottleneck is v.what.rast which converts
>>> the queried raster values into text with 15 decimal points?
>>> So is there any (future) way to have similar precision in a DCELL
>>> raster and its corresponding point vector (r.to.vect - v.what.rast)?
>>
>> The short answer is no because binary values have to be converted to
>> text which for fp values is lossy.
>>
> Thank you for the answer. Good to know.
>
>> The precision of double values in attribute tables is probably
>> dependent on the module loading values to the table and maybe also the
>> db driver.
>>
>>>
>>> Or is there any other recommendable way to get a "list" resp. loopable
>>> object that contains all raster cells (x,y and value) without loosing
>>> the precision of the original raster?
>>
>> Hmm. Something with r.mapcalc? Any *.[in|out].ascii will suffer from
>> the lossy conversion of binary -> text -> binary values.
>
> I also though about looping over the cells via accessing the raster
> as numpy array in python. Anyway I have to look into that to get
> the e.g. also x and y for each cell beside the value etc. I'll try it ...
>
>>
>> Also be aware of the difference between precision and accuracy. For
>> example, floating point DEM can have a precision in the nanometer
>> range, but its accuracy may be in the meter range.
>>
>> In your case, rescaling could help.
>
> Thank you for that explanation. In my case I am actually working with a
> "virtual probabilistic surface" rather than with real elevation values etc. Thus
> generally all values of my maps are between 0 and 1 with partly very small
> values. Of course in this case rescaling might help and I'll go for
> that approach for the moment.
>
> /johannes
>
>>
>> Markus
>>
>>>
>>> /Johannes
>>>
>>>> Markus M
>>>>
>>>>>
>>>>> Some additional info: I am working on Ubuntu 12.04 whit GRASS 6.5SVN. My
>>>>> point vector  is
>>>>> connected to an sqlite db - table (standard sqlite connection).
>>>>>
>>>>> Of course I thought about a work around (e.g. multiplication by a large
>>>>> scalar). Any other suggestions?
>>>>>
>>>>> Best regards,
>>>>>
>>>>> Johannes