I'm interested in a low-cost or open source solution for creating land cover GIS layers that utilize both spectral and textural extraction algorithms. I have used PCI Geomatica, ENVI, and Feature Analyst VLS in the past; however these solutions are a bit beyond my price range, any software recommendations?
Add a comment
|
3 Answers
You can use GRASS GIS for this which supports texture extraction and image classification based on a radiometric/segmentation approach. For an idea, check this conference abstract, a planned talk at the Geoinformatics FCE CTU 2011.
See also: http://grass.osgeo.org/wiki/Image_processing and http://grass.osgeo.org/wiki/Image_classification for an overview.
If I understand you correctly, you are looking for a supervised classification procedure. Some theoretical background: http://rst.gsfc.nasa.gov/Sect1/Sect1_17.html
This is certainly possible through grass: http://grass.osgeo.org/wiki/Image_classification#Supervised_classification_2
As an alternative you could also look at saga (I'm not saying it is better, I just know it better myself), which also plays nicely with qgis and R. There are some video's demonstrating this on this site: http://www.uni-koblenz-landau.de/landau/fb7/umweltwissenschaften/landscape-ecology/Teaching/geostat (download the datafiles to get the presentations).
In all gis programs, what you will do is define a number of reference points or polygons in one type of land, which are then extrapolated to the rest of the area. Here is an example of a landuse classification:
And in fact if you have drawn your training polygons in any gis program, you can use R to predict. Make an overlay with your grids, and then use any prediction system you like (eg rpart if you want classification trees). More info in this book around page 222: http://www.lulu.com/product/file-download/a-practical-guide-to-geostatistical-mapping/14938111
There is a lot more to say, you training sets should be representative for your study area (perhaps it would even be better to generate random points in R and to classify those). You should also choose your auxiliary datasets carefully, and you might want to generate new ones if eg texture is an important property.
--
If all you want to do is extract regions or features (without classifying them), a segmentation algorithm is more likely what you want. One example (implemented in SAGA GIS) is discussed in this paper: http://mirror.transact.net.au/pub/sourceforge/s/project/sa/saga-gis/SAGA%20-%20Documentation/GGA115/gga115_03.pdf
-
Thank you very much for your response. It seems you know exactly how one can accomplish my goals. What I would really appreciate is if you clarify your answer a bit more. I am specifically interested in the steps involved so that I can teach the program which features are correct and which ones are wrong until all (or most) of the correct features are extracted. Commented Sep 26, 2011 at 15:31
-
Provide more info (in your question, not in comments) what exactly the features are that you want to extract. Apart from that: if there is an overlap in the signal (see nasa link) of different land-use types (or whatever you are mapping), automated classification will not work well.– johanvdwCommented Sep 26, 2011 at 18:56
You would be able to do that with GRASS.
You will first work with raster data :
- I will point you out to that tutorial. See the raster part.
- You will use r.mapcalc and r.reclass to extract desired features.
- r.to.vect will permit you to vectorize your data.
Finally you will manipulate vector data. v.db.select and v.class will help you.
