--- Start of Idrisi List message
Hi everyone,
'Thank you very much' to all the people who took pains to reply and explain
the issues involved. I am grateful to them all. Here is a compilation of
the responses I got to my two postings. My first message was:
>In a trial resampling of Landsat TM image, I got follwoing results:
>Control points used in the transformation :
Old X Old Y New X New Y Residual
-----------------------------------------------------------------
50.955224 35.425512 687508.136072 5817365.827314 0.026984
46.999211 36.707037 678139.898016 5820095.309884 0.008801
56.657083 34.129128 701152.899997 5814602.226169 0.012568
62.979275 36.134994 716214.636796 5818867.042685 0.022410
80.066281 45.778009 756768.224205 5839116.391862 0.003913
11.612536 33.032220 594016.134369 5812414.807256 0.008329
13.681643 36.285292 599077.687102 5819311.753580 omitted
>Overall RMS = 0.016075
>Note : RMS Error is expressed in input image units.
Resolution is 30m.
My problem is how do I know if this RMS is acceptable/ significant?
----------------
Responses:
From; "Timothy M. Shields" <[EMAIL PROTECTED]>
Your rms looks great. You want it to be one half of your pixel resolution so
you have surpassed that. Pay attention to how many control points you need.
The mininum number depends on the mapping function and resampling type. It is
spelled out in the online help.
From: "Donald E. Simmons" <[EMAIL PROTECTED]>
There are some good articles on the quality of registration of grids;
however it has been several years since I have worked with this, so I don't
have a good or available list. The early documentation of ERDAS and
Arc/Info were very good descriptions of grid rms error. Geo-statistical
articles address this issue also.
Couple of points from my experience. When one does linear regression and
the data is a point one gets a very good r-square but the equation fails in
usage. Well the same happens with this type data. If the selected points
tend toward a straight line, one gets a good rms but the registration will
fail the more one moves at a right angle from the line. The registration
points need to define an area - square/rectangle. Do a scatter plot and
color code the plot with the rms value. low - green, some - blue, middle -
yellow, more - magenta, high - red. Look for the green blue as a straight
line and red at distance from line, ie poor registration. This is one item,
one topic but common and often overlooked. (I got a good rms, but it
doesn't seem to give good registration results, why?)
What I have done and this really fits the method and your data. Take groups
of 4 points and get the rms and equation; then use the other points to test
the quality of the equation. If a point often shows poor rms then the point
may be suspect. If different points show large rms then the total points
are weak, new points are needed.
From: Ron Glasson <[EMAIL PROTECTED]>
The real question is why are you resampling in the first place. The purpose
behind this will indicate the amount of error you accept. Is this for
georeferencing an image? or for changing pixel dimension to conform to
another data source? Your error tolerance can be thought of as a distance
from the original data source pixel or the source you are converting to. If
it is acceptable to transform the data within a given pixel radius then your
result is determined by an rms of 1. On the other hand if a distance is the
determining factor then in your case your error is within (30m x 0.016075)
= 48 cm
Have a look at an Erdas Imagine field guide if you have access to one (page
306)
From: "David R. Green" <[EMAIL PROTECTED]>
In part it will depend upon what you are planning to do with the image and how
important the RMS is to that task........................
---------------------------------------------
MY SECOND POSTING:
>Thank you very much for all the responses. The purpose of my resampling to
>detect change in tree cover on the basis of unsupervised classification
>(with ground truth derived from another previous study.
>
>In fact, the 'wonderful rms' value (0.016075) is the source of my worry. I
>do not know if I am interpreting it correctly. While resampling, I have
>used the procedure given in IDRISI manual for calculating no. of columns
>and rows, so a pixel represents 30m for the image. And I have done very
>well (?) by keeping rms within the maximum limit of 0.5 times the
>resolution. I never expected my resampling to be so 'perfect'. Is such
>result really achievable?
>
Responses:
From: rjs1 <[EMAIL PROTECTED]>
The RMS value that Idrisi gives is in the input units. By looking at
your data it seems that 1 unit in the old system is equal to about 2400
m. So, 0.016075 in the old units is an RMS value of about 38 m which
seems pretty good to me if your raster resolution is 30 m.
From: Ron Glasson <[EMAIL PROTECTED]>
I have yet to use idrisi for georeferencing but I have done it many times
(>4000 frames) in Erdas Imagine. The principal is the same but the procedure
for the two is somewhat different. The less control points down to the
required minimum of four (in imagine), improves the RMS but of course this
is a statistical aberration as it only calculates for those control points.
If you can achieve this RMS of 0.01 with 7 control points I suppose that the
flying height of the sensor is very high (hence a planer image in areas of
high relief) or the focal length is very long, or the terrain is very flat.
How big an area are you looking at? (just curious!!)
My question about why you are resampling did not relate so much to the
purpose of your study, but why resampling at all. Are you attempting to
match existing datasets pixel resolution? or are you attempting to remove
geographical distortion from an image. By resampling you need to be careful
that you don't remove actual data values through the process. Use a nearest
neighbour resampling regime rather than cubic convolution as the latter will
actually change data values which will become important later in your case
if your veg canopy is in proximity to water bodies or wetlands.
From: "Dr. Alberto Gomez" <[EMAIL PROTECTED]>
Hi there If you take a look under the help of Idrisi32 or Cartalinx You�ll
see a reference about the RMS which is acceptable. Even there is a lot of
info could be difficult to understand and apply in practice.
At the Institute of Geography (UNAM Mexico) we defined that an overall RMS
should not be greater than 1 pixel�s size whatever the units are.
From: Michele Fulk <[EMAIL PROTECTED]>
The RMS reported for raster resampling is in INPUT image units. If you set
your min/max x/y to be 0-number of columns and 0-number of rows, then you
multiply your RMS by the cell size to get the RMS in meters. Beware that the
fewer points you use, the easier it is to achieve a good RMS, but that
doesn't necessarily mean you have a good "fit". How many points did you use
and what mapping function?
From: Ed Wright <[EMAIL PROTECTED]>
I'm guessing from the numbers that the New X and New Y are UTM coordinates?
If that is the case, one "unit" in the input coordinates is approx 2380
meters.
- using the 5th and 6th Old X values: 80 - 11.6 = 68.4
- using the corresponding New X values: 756768 - 594016 = 162752 meters
- then: 162752 / 68.4 = 2379.4 (meters), that is each "unit" in the input
coordinates ~> 2380 meters
so your RMS error is .016 input units = .016 x 2380 ~> 38 meters
(But - I may have misinterpreted your coordinate systems)
In any case, that doesn't explain the very small RMS value. You don't say
what order polynomial you used for the mapping function. If you used a
quadratic polynomial, then 6 points is the minimum required (RMS should be
0). If you used a linear function, you have enough points but they are not
"well distributed". The points you have all fall close to one straight
line. With 6 points, you should strive to have one in each corner of your
study area, and one near the center.
-------------------------
Thanks.
vivek
--- End of Idrisi List message
IDRISI-L is an unmoderated list. As a service to list users, please post a
summary of useful responses to your questions.
To post a message to IDRISI-L, send it to [EMAIL PROTECTED] To
subscribe to the list, send an email from the address to be subscribed to
[EMAIL PROTECTED] with "subscribe" (without quotes) as the first line
of the message body. To unsubscribe, do the same, only with "unsubscribe"
as the first line of the message body.
