Top Document: Satellite Imagery FAQ - 3/5 Previous Document: How good are classification results in practice? See reader questions & answers on this topic! - Help others by sharing your knowledge 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 ******************************************************************** User Contributions:Comment about this article, ask questions, or add new information about this topic:Top Document: Satellite Imagery FAQ - 3/5 Previous Document: How good are classification results in practice? Part1 - Part2 - Part3 - Part4 - Part5 - Single Page [ Usenet FAQs | Web FAQs | Documents | RFC Index ] Send corrections/additions to the FAQ Maintainer: satfaq@pobox.com
Last Update March 27 2014 @ 02:12 PM
|
Russians who want to impress their social media followers can now rent huge bouquets of flowers for just long enough to snap an Instagram worthy photo, It's revealed.
Pop up services are advertising on social networks in planning for International Women's Day on 8 March, A public holiday in Russia when women are in the past feted with flowers and other gifts. They're offering 10 minutes with an enormous bouquet sufficient time to perfect the best angle and pose before the courier takes it back, The TJournal news blog reports. One account offers women a fleeting visit from 101 roses for 700 roubles ($12; 10).
Posting selfies with huge bouquets seemingly sent by a boyfriend or secret admirer has been a trend among Russian Instagram and VKontakte social network users for quite a while.
The 360 TV website contacted the master of one flower rental Instagram account, [url=https://charmdatescamreviews.wordpress.com/tag/hot-russian-women/]russian sexy girls[/url] Who insisted his service was real and had received many asks. He says the 10 minutes can be drawn out a little, But not consistently. "the most important thing to avoid is: 'I'm obtaining put some make up on and tidy up' and all that', according to him.
Another account that 360 TV contacted ended up being a joke set up to mock the new trend, And there was plenty of ridicule from Russians placing comments online. "the secret of a mysterious bunch of flowers from a stranger has been EXPOSED, Writes an individual on Twitter. "might be quintessence of modern values, Says an Instagram surfer, Who suggests similar service aimed at men: "A Rolex download, Two iPhones shared and a selfie taken in a Moscow office with a panoramic view.