Model Benchmarking

[Keven Bennett]

    Figuring out what is 'reasonable' is another vagueness: statistically
speaking, 95% confidence would probably be enough, and I guess this means doing
the sampling you suggest in numerous selected areas and get a statistacally sound
result, and an error bar for 95% (I forgot how many STTDEV's away from the mean it
is for a two tailed test, I can look that up).

    As far as AV, then, I do lose the polys w/in cell boundaries with less than 50
sq. meters?
    Keven



Phil Hurvitz wrote:

> Hi Keven, what is considered 'reasonable?' I'd take the approach of using
> a smaller section of the study are and using smaller cell size, then
> progressively moving up in cell size.  You might get a feel for a minimum
> size to properly model the phenomenon.  If you can answer the question is
> "what is the smallest patch of spartina that can be considered a source
> for spread?"  That might drive your initial choice of cell size.
>
> The problem with AV-SA's rasterization
>
>         anFTab.AsGrid ( aPrj, aField, gridSize)
>
> is that it doesn't give you any choice on what to do with partial cells,
> whereas ArcInfo uses
>
>         POLYGRID <in_cover> <out_grid> {value_item} {lookup_table}
>         {weight_table}"
>
> the weight table lets you choose how to handle partial cells).  So you are
> stuck with small polys that disappear rather than creating single cells,
> if that is what you wanted to do.  I'm sure other spatial modeling
> software gives you this option as well.
>
> -P.
>
> ******************************************************************************
> Phil Hurvitz, MFR | GIS Specialist | College of Forest Resources | 355 Bloedel
> Box 352100 | University of Washington, Seattle, Washington  98195-2100, USA
>  tel: 206.685.8179 | FAX: 206.685.3091 | e-mail: phurvitz at u.washington.edu
>              WWW: http://lobo.cfr.washington.edu/~phurvitz/
> ******************************************************************************
>
> On Mon, 22 Jan 2001, Keven Bennett wrote:
>
> >     I am working with ONRC in developing an iterative spread model of
> > Spartina in Willapa bay.  The intent of the model is to give one the
> > ability to 'see' how Spartina is distributed on the ground in a given
> > time in a given year at 10 meter resolution.
> >     We have the model running with perameters we believe are reasonable
> > representations of the primary factors that affect the spread (with the
> > exception of magnitude and direction of streamflows).
> >     We want to benchmark our output with 1994 and 1997 data from DNR
> > which is fairly accurate high resolution vector data derived from CIR
> > aerial photography, but we have issues to address first:
> >     We can't directly compare the two years with the two years
> > corresponding output because the vector data captures polygons with
> > areas much smaller than 100 sq. meters (10 meter cells).  The resonable
> > assumption is that the model output should have less total area than the
> > '94 and '97 data we want to accept as true.  The question is how much?
> > Any suggestions on arriving at a figure for this?
> >     If it is safe to make the assumption that any polygon less than 50
> > sq. meters using ArcView's rasterization algorithms (convert to grid in
> > correct analysis window), then can we justify that a match of our model
> > output to the area arrived at by rasterizing the DNR data as 'reasonably
> > accurate'?
> >     These are kind of vague questions, but this is the kind of issues we
> > are confronted with.
> >     Keven Bennett
> >
> >
> >