The aim of this blog is to encourage the exchange of ideas between remote sensing and GIS users, particularly those interested in image processing, image transforms, LiDAR data, hyperspectral imagery and interferometric DEM processing.
Readers of CPRSI are especially welcome - report typos, misspellings and errors.
I am especially interested in comments from bloggers who have tried to use the MIPS package that can be downloaded from:
http://www.nottingham.ac.uk/~lgapm/Computer-Processing-4.html
You can also find Powerpoint slides of all the diagrams in the book (in full colour, of course) as well as four worked examples with full datasets.
The book is available from Amazon. The Amazon UK URL is:
http://www.amazon.co.uk/Computer-Processing-Remotely-Sensed-Images-Introduction/dp/0470742399
To view comments or to add a comment of your own just click on CPRSI under ARCHIVE in the right hand column.
Good wishes to all.
I saw this and it has made me think of the distances in our minds between different parts of ourselves.I suppose one could say the inner geography of our souls.Perhaps not geo as that relates to earth,but the inner aspect of that.It may be good for a poem though that is not what you are expecting from blog readers
ReplyDeleteThere is a world inside us too.So maybe my comment makes sense?
ReplyDeleteCLASSIFICATION, SVM, SAMPLE SIZE
ReplyDeleteMy name is Rener, I am master student in Forestry and Environmental Sciences at the University Federal of Mato Grosso, Brazil. I'm working on my MA thesis with the mapping of wetlands in the Araguaia region-Bananal Island, here in Brazil. I have to generate vegetation maps of my area of study, and I am suing my images with the SVM, and I'm having some doubts about the digital image processing. I've asked several people, and still can not reliable answers about it. If you can answer me, I would be very grateful.
1) I wonder what the minimum number of training samples for a given class to make a map of vegetation types? There is some literature to tell me this because I researched and found no research on this.
2) I would also like to know if there is a specific pattern to the size of training samples (eg.: Number of pixels per sample). This varies from classifier to classifier? For the SVM classifier which would be the ideal number of training samples and what the optimal size of these samples in number of pixels?
3) Another question is in relation to the field to be used as an aid in both sampling and validation. For example: If I have 150 points of ground truth, how many points I can use in sampling and how do I use in my validation. For I believe that I can not use both at the same points as in the validation sample, for my statistical analysis would be biased. There are also some literature about it? Or do I just use the points collected in the field validation?
CLASSIFICATION, SVM, SAMPLE SIZE
ReplyDelete1. If you are using SVM one of the major advantages over statistical methods is that only those training pixels lying near the boundary between two classes are used. See Mather and Koch, 2011, Figure 8.21. page 267.
2. For the statistical ML classifier I recommend 30 samples per feature. Anns seem to manage with smaller sample numbers as the ML method has to estimate a variance-covariance matrix for each class whereas the ANN does not. SVM should manage with even fewer samples than an ANN, as stated in the answer to Q1 above. See Mather & Koch, 2011, page 84, left hand column.
3. You are right - you should not use the same data for setting up the classifier and testing it. The training and test data should be kept separate. The more test data you use the more rigorous are your results (as you could show by computing the kappa coefficient and its standard deviation. You really need to consult a good book on sample design. There used to be a good one by Kendall which described all sorts of things such as sampling with and without replacement as well as sample design (e.g., a spatially randomised sample is better than a systematic grid sample). The idea is to represent adequately (i) the mean location of the centre of the class in feature space and (ii) the spectral variation around this mean point. You could consider putting your test data through a screening programme as described by Mather and Koch, 2011, page 243, Campbell (1980) method.
My name is Atthur, I am master student at the Geography Faculty of Gadjah Mada, Yogyakarta, Indonesia.
ReplyDeleteI am currently working on a thesis : dry land interpretation of image
data ALOS AVNIR-2 with using Tasseled Cap Transformation(TCT).
Do you know the transformation coefficients for image ALOS AVNIR-2 TCT?
Dear Atthur
ReplyDeleteI cannot find any reference on Google to the tasselled cap transform coefficients for ALOS AVNIR-2. However, if you look at CPRSD page 160 you will see reference to a paper by Jackson (1983) who describes a method of calculating TCT coefficents for any sensor data. Jackson's method is implemented in the MIPS software which you can download free from http://www.nottingham.ac.uk/~lgapm/Computer-Processing-4.html If you are using time series of TCT images you should atmospherically correct each image set and convert to radiance. Otherwise athmospheric effecs will mask any real changes.
Regards
Paul Mather
MIPS UNDER WINDOWS 7
ReplyDelete--------------------
I haven't had any reports of problems running MIPS under Windows 7 but to be on the safe side try right clicking the desktop icon for MIPS, select Properties then the Compatibility tab. Set compatibility mode to Windows XP (Service Pack 2) and you should be safe.
Good Luck
Paul Mather
Google Chrome does not properly display the web page at http://www.nottingham.ac.uk/~lgapm/Computer-Processing-4.html. I cannot see the reason for this at the moment. The short-term solution, until I discover the error (which may be in Chrome!) use Internet Explorer or Firefox, both of which run OK.
ReplyDeleteThe web page now works with Google Chrome browsers.
ReplyDeleteFor some reason the exercises zip file was not being found by the browser. This has now been fixed. Feedback from users of the exercises would be useful.
ReplyDeleteHi Paul,
ReplyDeleteThe web page hosting the code is not accessible by firefox, nor by google chrome. This is the displayed message upon attempt to connect:
Forbidden
You don't have permission to access /~lgapm/Computer-Processing-4.html on this server.
Could you look int othis?
Regards,
Max
April 2015
Hi Paul,
ReplyDeleteI am having the same issue as Max aka Joker.
403 Forbidden
You don't have permission to access /~lgapm/Computer-Processing-4.html on this server.
Apache/2.2.14 (Unix) mod_ssl/2.2.14 OpenSSL/1.0.0g Server at www.nottingham.ac.uk Port 80
Regards
Chris
hello
ReplyDeletei faced with the same problem.
i don't have permission to access /~lgapm/Computer-Processing-4.html on this server.