Hi Amy, You can also use .erv files as your reference/ground truth dataset (does not have to be ascii). One way that might work is to display your ground truth data in ER Mapper and save it as .erv using the annotation system.
Have included some info from an old email on this which may be of some use... Good luck. cheers Mike > -----Original Message----- > From: Manuel Alonso [mailto:[EMAIL PROTECTED]] > Sent: Tuesday, 12 September 2000 6:32 PM > To: '[EMAIL PROTECTED]' > Subject: RE: Confusion Matrices > > > Neil: > > You can "crosstab" three kind of datasets against the classified image by > using three methods (although the most common one is 2): > > 1.- A vector file with points and associated attributes: this is based in > the fact that usually the "field truth" is basically a point that people > register in the field (that is, an idependent sample, to test how the > classification has gone; some authors think that using the training fields > as the "truth" is not statistically reliable). You can build this sort of > file in ER Mapper simply importing an "ASCII points with attributes" file > (Utilities - Import vector and GIS formats - ASCII points with attributes). > That import will create a point file in .erv format where each one of the > present points is "on" a pixel in the image, and where that point has got an > associated attribute, that is the number of one of the classes in the > classified file you want to verify. > > 2.- Using regions: there is another method - possibly the easier and most > used one - to "crosstab" two datasets for making the confussion matrix, and > it is to use "vector regions" inside any file to use them as the "truth". If > you've got some regions defined in any raster dataset, you can use that > dataset as the "reference dataset" to cross against the classified image, > but to work properly the names and numbers of the regions should match with > the names and numbers of the classes in the classified dataset. This is > quite "normal" when the user wants to check his supervised classification. > In that case, the classified image is "crossed" against the ORIGINAL dataset > used to classify the data (reference dataset), since the vector regions are > the "training fields" defined in that original dataset. Since the "training > fields" are suppoused to be "true", the result of this operation is an > estimation about the degree of accuracy of the classification with respect > to the areas used as "training fields" (this answers the question "how much > of the area covered by my training fields for Class A has been classified as > Class A?"). Notice that this sort of check is the most common one (and easy > to do), since accurate classified maps from the area to classify are not > common. However, the statistical significance of this sort of check depends > on the number of training fields, as you probably are aware. > > 3.- A raster file that contains a classified image inside. Obviously, to > make this work, it is necessary to have an already classified image with the > "ground truth" inside (for instance, a map of the same area), and that is > not usual. But sometimes, we want to validate how well a classification > works on some data for some "well-known" classes, to extrapolate the method > to areas "unknown"; for instance, imagine you want to check how good is a > Landsat 7 image to try to detect lithology: you classify the image and you > want to check it against an existing accurate geological map. > > The main question here is how to create that "truth map" raster image from a > simple raster image (for instance, a scanned map). As I told you before, ER > Mapper confussion matrix system was designed to use a classified image or an > image with vector regions (see 2) as the reference dataset. As you are > aware, that sort of map should have the Digital Numbers (values of the > pixels) assigned to the same numbers used in the classified image you want > to verify. For instance, imagine a classified image with the following > categories: 1 bare soil for cultivation, 2 sparse vegetation, 3 vegetation > (irrigated crops), 4 sandstone and 5 is sand. Now, imagine you.ve got a map > to verify the truth of that classification. That map should have at least > the same categories (the same numbers assigned to the pixels) than the > classified image. To "build" that map, the header of the ER Mapper image > should content these blocks: > > BandId Begin > Value = "Classified" > BandId End > > In a simple raster image, this block value is "Pseudo layer", or any other > name. > > After that block, you'll find all the classes defined (one per Digital > Number present in the image. In this case, 5): > > RegionInfo Begin > Type = Class > RegionName = "Bare soil for cultivation" > ClassNumber = 1 > RegionInfo End > RegionInfo Begin > Type = Class > RegionName = "Sparse vegetation" > ClassNumber = 2 > RegionInfo End > RegionInfo Begin > Type = Class > RegionName = "Vegetation (irigated crops)" > ClassNumber = 3 > RegionInfo End > RegionInfo Begin > Type = Class > RegionName = "Sandstone" > ClassNumber = 4 > RegionInfo End > RegionInfo Begin > Type = Class > RegionName = "Sand" > ClassNumber = 5 > RegionInfo End > > If you want to convert any image in this sort of "classified image" to use > it as the "truth" for the confussion matrix, you should modify the header > file of the ER Mapper image by following these guidelines. To modify the > header file, just edit the .ers file with any text editor (as Wordpad). > Again, for this to work fine, the Digital Numbers of the image should match > with the numbers assigned to the classes. > > Regards, > > Manuel Alonso > Earth Resource Mapping > Blenheim House > Crabtree Office Village > Eversley Way, Egham > Surrey, TW20 8RY, UK > Telephone: +44 1784 430-691 > Facsimile: +44 1784 430-692 > > > > -----Original Message----- > From: Neil Sims [mailto:[EMAIL PROTECTED]] > Sent: 12 September 2000 01:06 > To: ER Mapper List > Subject: Confusion Matrices > > > Hi People > > I posted this a little while ago but got no reply. I'm sorry to have to ask > this question again. Its probably something I am overlooking in the on-line > manuals but... > > I was wondering if someone could point me in the right direction to get > information about the confusion matrix facility in ER Mapper v 6.1. I need > to know about how to set up my reference dataset so that individual class > and overall accuracy assessments are performed correctly. > > Any help would be great. > > Thanks. Neil. > > > **************************************** > Neil Sims > > PhD Candidate > University of Canberra, Australia > Phone: +61 (02) 9818 4770 > email: [EMAIL PROTECTED] > > "Resourcefulness is next to Intelligenceness" > **************************************** > > -----Original Message----- > From: [EMAIL PROTECTED] > [mailto:[EMAIL PROTECTED]]On Behalf Of Amy Keister > <[EMAIL PROTECTED]> > Sent: Thursday, 19 September 2002 1:30 AM > To: [EMAIL PROTECTED] > Subject: er mapper and arcgis > > > Hello, > > I have my ground truth data saved in ArcGIS as a coverage. I have queried out > the portions that I want to use for accuracy assessment, but I can't figure > out how to import it into ER Mapper. The help screen says that I need an > ASCII format with points and attributes. The import utilites in ER Mapper > will only import lines and polys from a coverage and the export utility in > ArcTool box doesn't seem to have that option. > > Any help would be greatly appreciated, > > Thank you, > > Amy Keister > Graduate student > Southwest Missouri State University > [EMAIL PROTECTED] > > ----------------------------------------------------------- > > To make changes to your subscription, please visit our website, >http://www.ermapper.com/technicl/ermapperl/index.htm > > ----------------------------------------------------------- To make changes to your subscription, please visit our website, http://www.ermapper.com/technicl/ermapperl/index.htm
