I need to classify a mosaic of several images. How best to do it?
  
   David Schaub (dschaub@dconcepts.com) posted a question on this. Here
   is his summary of replies:

Dear Netters,

Some time ago I posed a  question to this list with regards to classification,
rectification, and mosaicking.   My original question was as follows--

>Hello,

>We need to georectify, mosaic, and classify several (3 or 4) Landsat TM
>scenes using ERDAS Imagine.  The classification will need to show major
>land cover categories, such as bare ground, grassland, shrubby range,
>built-up, coniferous forest, broad-leaf forest, water, etc.  In the past
>when we have done this the seams between images are quite evident in the
>classification.  We would like to minimize differences between images, yet
>be asaccurate as possible in the classification of each image.

>My main questions are these -- Should we classify each image separately
>and then mosaic them, or should we mosaic the images first and then
>classify them?  Can georectifying the images effect the classification?

>You can assume that images along a path will have the same acquisition date,
>however scenes on adjacent paths will have different dates (at least by two
>weeks).  I will post a summary.  Thanks in advance for your opinions :-)

This quickly generated a flood of responses.  While there wasn't complete
agreement, the majority of respondents believed that I should first classify
the images, then do the rectification and mosaicking.  Nearest neighbor
should be used when rectifying the classified image (or if the image data
are rectified before classification).  Thanks to all who responded!!  Comments
are summarized below:

David Schaub
dschaub@dconcepts.com

*******************************************************************

     I have done the same things you are attempting to do for my thesis work.
I think the best course of action would be to classify the images first, then
rectify the images and then merge or mosaic the images.  Rectifying the images
before you classify may distort the spectral characteristics of pixels and
thereby influence your classification.  Furthermore, the smaller the area you
are classifying, the more accurate the classification will be, so if you
mosaic a large area and then attempt to classify the mosaiced image, there
will be more confusion possible based on the heterogeneity of a larger area. I
hope this helps, contact me if i can be of further assistance.

David Smith

*************************************************************

Here's my 2c for what it's worth...

I classify TM scenes separately and then mosaic the classifications. My
classifications almost never have a seam in them...If there is a seam
it is usually due to the difference in the date of the scene. You have
to be careful though... you need to use the same method of classification
(plotting out feature spaces and elipses helps) for overlapping scenes.
Sometimes this is why people use the other method...

If you're going to do this the other way round...i.e. mosaic and then classify
scenes you will have to calibrate the scenes to radiance and then use some
kind of atmospheric correction before mosaicking them.  This should in theory
minimize the difference in the spectral information between scenes....I would
avoid using any kind of histogram equalization ...although it may look nice,
you are loosing the original pixel information.

\\.   _\\\_____
 \\\ /ccccccc x\        Fiona Renton, GIS and remote sensing analyst
  >>Xccccccc(   <       CALMIT, Conservation and Survey Division
 /// \ccccccc\_/        University of Nebraska-Lincoln
''     ~~~~             renton@fish.unl.edu

**********************************************************************

What sort of classification?   Pixels?   Clusters?   Polygons?
Higher-level features?    If your classification units are homogenous
and shape is not important, you should clearly do it before mosaicing.
If not, you have a genuinely interesting problem, and will probably have
to your own research (starting at your local academic library, assuming
there is one :-)

Nick.

*********************************************************************

Geo-rectification will have a small effect on classification due to the
resampling process.  I can't help to much on classification part,
because that is not my area, but my feeling is that mosaicking
non-classified images may be easier than trying to match features in a
classified image.

Ok, this is my area.  You can not assume that images on the same path
are imaged on the same day, However, they could be.  You should be able
to check the meta-data to find out if they were.  The next path west
could have been imaged 7 days after the path of interest or 9 days
before and the next path east could have been imaged 9 days after the
path of interest or 7 days before, again check the meta-data.  The next
chances are to add 16 days on to those numbers (i.e. 7 + 16).

This is true for Landsat 4 and 5 only (will be true for Landsat 7).

Chuck
wivell@edcsnw38.cr.usgs.gov

************************************************************

Yes the georectification process will affect the classification
results.  My suggestion is to classify each individual image first and
then mosaic them together.  I have done this before and it works well.
If you mosaic first and then classify you have to calibrate the data,
apply radiometric corrections etc...  Not worth the trouble in my
opinion, and you probably won't get any good results.

The resampling technique (convolution) will affect the radiometric value
of the image and may not be suitable for adequate identification
aftrewards.  To avoid visible seams, just go around the areas, try to
contour the natural groupings (classes after classification)


To resume, in my opinion, if you want good accurate results: Classify first
and after mosaic.

Francois Beaulieu

************************************************************

You definitely want to mosiac the 4 images first (into one file) and
then run the classification on that.  Because of subtle differences in
the radiometric characteristics of each image, the classes in separate
classifications will rarely "line up" perfectly when mosaicked afterward.

> Can georectifying the images effect the classification?

Yes it can, depending on the resampling technique you use.  When
rectifying the images, use Nearest Neighbor resampling as that
will ensure that original pixel values are used to create the
new rectified dataset.  (Bilinear or Cubic will average the
original data, resulting in slight degradation.)  I would:

1)      Rectify the four images (use Nearest Neighbor)

2)      Contrast balance them, using for example Histogram
        Matching or another technique.

3)      Mosaic the four contrast balanced scenes into one file.

4)      Run the classification.

I hope this helps.


Eric Augenstein
Manager of Training Services

*************************************************************

In general you can't depend on the DN values from one image to the next
to be related. You should classify before your mosaic - in other words
mosaic the classification, not the images. Otherwise you mix unrelated
DN values into a signal classification which would be wrong.

Classification may be affected by geo-rectification. If the
geo-rectified image has the same pixels and pixel values as the
original, the classification should not be affected. However, this is
an unreal assumption. A geo-rectified image will almost always have
resampling - which means that pixels are either dropped or replicated -
unless a filter is applied (like bilinear or cubic convolution) in
which case the pixel values change as well.

If the classifier is single pixel based (like isodata) then the
classification is only affected by the resampling as the sigatures are
affected by the replication or dropping of values. If the classifier is
regional or global (like multi-resolution/multi-scale classifiers, or
region linking) then the classifiers may be affected to a greater
degree.

You can classify before or after geo-rectifiction and the results
will not be vastly different.

But the bottom line to mosaic at the very end.

Michael Shapiro                   mshapiro@ncsa.uiuc.edu
NCSA                              (217) 244-6642
605 E Springfield Ave. RM 152CAB  fax: (217) 333-5973
Champaign, IL 61820

********************************************************

Re Michael Shapiro's posting,

There is no doubt that that you cannot depend on the DN values from one
image to the next (especially with images from adjacent paths which are
taken on different dates (see Chuck Wivell's posting).

However mosaicing images which have been classified seperately may produce
unusual results ie trying to match classes from different images.

A suggestion would be to first try some kind of atmospheric correction on
the images, mosaic them and then classify them together. Assuming

i) you can do a credible atmospheric condition (using Dark Pixel
Substraction, Band Regression etc) plus, perhaps, correct the images to a
constant solar elevation angle
ii) the images from different paths were not taken on widely different
dates and
iii) (linked to ii) the ground conditions are similar for the images
from different paths

then the DN values between images should be comparable.

Euan

************************************************************

We are currently doing a statewide land cover classification for Mississippi
using TM scenes (10 of them). My responses for your questions:

   1. We classified each scene separately - mainly because the dates differed
      and in the cases where we had adjoining scenes taken on the same day, it
      was decided that classifying a full scene was a big enough task in both
      computer and human resources. If you had subscenes, it would not be too
      bad. I would advise against mosaicking scenes before classifying - your
      signatures for the same landcover class in the other scene(s) would be
      different and it would be a nightmare. Matching techniques that changed
      image pixel values would change your original data and corrupt your
      classification.

   2. We also georeferenced each scene before classification for the following
      reasons:

        - georeferenced ancillary data sources (roads, streams, NWI, etc) were
          used - including leaf-off TM scenes already in-house.
        - the need to have maps to take into the field for pre and post
          classification checks.

      We used nearest neighbor. This doesn't change pixel values but just
      moves them to a different location. In our case the image statistics
      were unchanged after georectification although it is probable that some
      pixels may be dropped or replicated (but when you georeference the
      classified image, those same pixels are going to be affected anyway).

Bottom line would be to classify each scene separately.  I would georeference
each TM scene first - when the classiciations are completed, stitching is
easy.

Jim

************************************************************

Our lab has had luck using regression techniques to mosaic the three
bands together.  Using ERDAS imagine, the steps are:
    1)  create an image where the two scenes overlap (this is best
            done with modeller, not layerstack:  layerstack only
            uses the geographical boundaries, whereas you want to
            have the area where there are values in both images
    2)  Use the Accuracy Assessment module to create random points
            on the image and remove those points which lie in cloud
            or shadow.
    3)  Export the X,Y coordinates from the random points and use
            these as a point file in the Pixel-to-Table function.
            Use the overlap image as the output image (make sure
            you have all the bands you want to regress (ie. image
            one's band 3,4,5 on top of image 2's 3,4,5
    4)  You now have a set of points that can be imported into
            any standard statistical package.  You need to have the
            values from the "larger" or primary image be the Y values
            and the other image be the X value (I'm told the correct
            statistical term is that the Y is the master and the X
            is the slave).

This should create a seamless image.  Obviously, the closer the B
number in the Y= bx + constant equation is to 1, the less you are
transforming the values of your slave image. We have also tried doing
classifications of each image first, but the results have been
disappointing.

Regards,
Sean Murphy
University of Maine

********************************************************************