| Re: Multi-band sampling strategy [message #72169] |
Wed, 18 August 2010 14:11  |
Maxwell Peck
Messages: 61
Registered: February 2010
|
Member |
|
|
On Aug 19, 6:58 am, JessW <jess.wal...@gmail.com> wrote:
> Hi Allard,
>
> Interesting. You are correct; I'm using random pixel locations across
> an entire Landsat scene, so many of the pixels may fall in the same
> slice. I had assumed that since I specify a unique line/column
> address in the ENVI_GET_SLICE command, IDL retrieves only the values
> for that one pixel. But obviously if the command retrieves the entire
> slice prior to defining the column selection, it makes for an
> inefficient process, as you point out.
>
> I've tangled with histograms and reverse indices before; the
> experience left me with a splitting headache, so it may take me some
> time to successfully apply your suggestion to my program. I'll post a
> follow-up message when I have results.
>
> Thanks very much for taking the time to reply!
>
> Jess
Jess,
Allard's approach is certainly better. I was unclear initially if you
were subsetting 'rectangles' or random point locations. As Allard
suggests point ROI's can get slow if you have a lot.. It all really
depends on how many points we're talking.
Max
|
|
|
|
|
|
| Re: Multi-band sampling strategy [message #72171 is a reply to message #72170] |
Wed, 18 August 2010 12:46  |
wita
Messages: 43
Registered: January 2005
|
Member |
|
|
Dear Jess,
The procedure that you are now using is fetching a complete slice of
your image
for each sample location. I do not know how large your image is, but
if I take
a typical TM scene (6000x6000 pixels), then many of your 38,000 pixels
will
actually be positioned in the same slice. So you are probably fetching
the
same slice a couple of times from your image just to take out another
sample
location. The more sample locations you have, the worse the speed
penalty gets.
Instead, what you could do to speed this up is first sort out which
sample
locations fall on which slice and group them together. There is
actually a
very efficient way to accomplish this in IDL by using the Histogram()
command
in combination with the REVERSE_INDICES. If you have never heard of
this then
please have a look at the histogram tutorial on David Fannings site to
understand the trick.
So in pseudo code it could look like this, where I assume that
SampleLines,
SampleCols store the line/column numbers of your sample positions:
h = Histogram(SampleLines, MIN=0, REVERSE_INDICES=ri)
FOR j=0, N_Elements(h)-1 DO BEGIN
IF ri[j+1] GT ri[j] THEN BEGIN
SampleColsinLine = samplecols[ri[ri[j]:ri[j+1]-1]]
SampledSpectra = Get_Sample_Spectra(fid, j, SampleColsinLine)
ENDIF
ENDFOR
with the function Get_Sample_Spectra() looking like this:
FUNCTION get_sample_spectra, fid, line, samplecols
Slice = ENVI_Get_Slice(FID=fid,line=line, pos=bandsSelected)
RETURN, Slice[samplecols, *]
END
This has the additional advantage that you are progressing through the
file
in the order in which it is stored on disk. This automatically takes
advantage
of the read-ahead capabilities of operating systems and disk drives.
Creating point ROIs as Maxwell suggest is, in my experience, also
becoming
very slow for a large number of samples. The suggestion by Paulo is
inherently
limited by the size of your memory. Moreover, it is very inefficient
for small
sample sizes.
Hope this helps.
Allard
|
|
|
|
| Re: Multi-band sampling strategy [message #72173 is a reply to message #72171] |
Wed, 18 August 2010 11:09 |
penteado
Messages: 866
Registered: February 2018
|
Senior Member
Administrator |
|
|
On Aug 18, 2:26 pm, JessW <jess.wal...@gmail.com> wrote:
> Hi Max,
>
> Thanks for responding, and sorry about the fuzzy description of what
> I'm doing. Let me try to clarify.
>
> I am tracking the reflectance change of particular pixels across a
> time-series of Landsat data. So for my current stack of 8 Landsat
> scenes spanning April - October, the program returns 8 reflectance
> values for a given pixel address x,y. ENVI_GET_SLICE seemed ideal for
> this task because it can sample a single pixel location through a
> number of bands; however, as I mentioned, it is very slow when it
> loops over a large number of pixels to sample.
>
> ENVI_GET_DATA appears to work on only one band, so it doesn't seem
> quite as suitable for what I'm trying to do. I'll have to investigate
> the use of ENVI_CREATE_ROI; thanks for that suggestion.
>
> Jess
If your image is not too large to fit in memory, it seems that it
might be better to just get the whole image as a 3D array, once, then
take out the subset you want directly, just with indexing. If the
image is too large, but the section you want fits into a small
rectangle, you could retrieve just that region, defined by the max and
min of your indexes, as a 3D array, then pick the points inside it.
|
|
|
|
|
|
| Re: Multi-band sampling strategy [message #72183 is a reply to message #72174] |
Tue, 17 August 2010 23:35 |
Maxwell Peck
Messages: 61
Registered: February 2010
|
Member |
|
|
On Aug 18, 1:10 pm, JessW <jess.wal...@gmail.com> wrote:
> Hi all,
>
> I'm using an IDL script in ENVI to retrieve the values of a multiband
> image at a user-specified number of pixels. The script works as
> intended, but is quite slow; sampling an 8-band image at ~ 38,000
> pixel locations takes
> 52.35 minutes. The issue is the following loop:
>
> FOR i= 0L, nSamplePix - 1 DO BEGIN
> pixValues[i,*] = ENVI_GET_SLICE(fid=fid,line=(index[1,i]-1),$
> pos=bandsSelected,xs=(index[0,i]-1), xe=(index[0,i]-1))
> ENDFOR
>
> If anyone can suggest an alternate, faster sampling strategy--
> preferably one that doesn't include the purchase of a new, faster
> computer--I would be extremely grateful.
>
> Thanks,
>
> Jess
I'm not clear on what you're doing but have you tried envi_get_data
with the dims keyword, or if that is not appropriate maybe creating a
point or polygon ROI (ENVI_CREATE_ROI/ENVI_GET_ROI_DATA)?
Max
|
|
|
|
|
|
|
|
comp.lang.idl-pvwave archive
Members
Search
Help
Login
Home
