[GRASS-dev] G7: r.neighbors changes data type from CELL to DCELL

[Markus Neteler]

On Thu, Apr 10, 2014 at 8:56 PM, Glynn Clements
<glynn at gclements.plus.com> wrote:
>
> Markus Neteler wrote:
>
>> in r59668 and r59669 I have fixed the issue that a map type
>> (CELL/FCELL/DCELL) was not preserved as expected but always converted
>> to DCELL.
>>
>> Now CELL is generated for
>>
>>     {c_count, w_count, NO_CATS, 0, 0, 1, "count", "count of non-NULL values"},
>>     {c_divr, NULL, divr_cats, 0, 0, 1, "diversity",
>>     {c_intr, NULL, intr_cats, 0, 0, 1, "interspersion",
>
> Note that this change is incorrect for method=count when used with
> weight=, as the result is the sum of the weights, which isn't
> necessarily an integer.
>
> Fixed in trunk in r59677.

Thanks,

>> Remaining issues:
>>
>> 1) "quantile=0.5" creeps into the history:
>>
>> r.info landuse96_28m_mode7x7
>> ...
>>  |    r.neighbors input="landuse96_28m" output="landuse96_28m_mode7x7" met\   |
>>  |    hod="mode" size=3 quantile=0.5
>>
>> --> quantile=0.5 should not be there (even if ignored).
>
> Remove the line:
>
>     parm.quantile->answer = "0.5";
>
> Optionally change:
>
>         out->quantile = (parm.quantile->answer && parm.quantile->answers[i])
>             ? atof(parm.quantile->answers[i])
>             : 0;
>
> to default to 0.5 rather than 0.
>
> But given that we already a distinct "median" method, using
> "quantile=0.5" is largely redundant.

... mmh, not sure what to change.


>> 2) The method=mode does not preserve CELL (likewise in r.series).
>>
>> Any suggestion how to fix these two remaining problems?
>
> You could just set is_int=1 (although that would make r59677
> incorrect, as the weighted mode calulation does return an integer
> result for integer values).
>
> Using method=mode isn't particularly meaningful on floating-point
> data, as the result is an artifact of the quantisation of the original
> data. If measurement was perfectly accurate, every value should be
> unique. This is largely why r.stats.zonal (formerly r.statistics2)
> doesn't have method=mode.
>
> Alternatively, replace the boolean is_int field with either a pointer
> to a function which returns the result type based upon the input type
> and whether weights are being used, or an enumeration (which would
> either select such a function, or an equivalent block of code within a
> switch statement).

This is a bit over my programming skills.
But the broken "mode" operator is a critical issue.

Help (=svn commits) welcome,

thanks
Markus