Hi Curt! Curtis Olson wrote: > Part of the data is based on a "new" method of automatic landcover > classification from Landsat satellite imagery. The method is not "new" > in that it is well-known in other areas. The method is "new" in the > sense, that it was not yet applied to FlightGear scenery generation. > > > Hi Ralf, this sounds very exciting. Is it something you are running > locally, or part of a larger external project somewhere? Can this > process locate lakes and rivers with any level of accuracy? What image > resolution is available. At some point it would be fun to experiment > with drawing the textures directly over the terrain ... as an option for > people that like blurry airports and taxiways that disappear into mush > when you get close to them. :-)
This is currently a local project. I am manually fetching the respective Landsat tiles (ETM+, 8 channels) and do manual training by marking some representative areas of different types. The goal is - as I said - to integrate this with OSGeo, who are also interested in the resulting data, and to use such data for the whole world to replace the polygonal features of VMAP0. There is a European project called CORINE, and they were obviously able to distinguish over 40 different classes of landuse from Landsat imagery by automatic classification. See http://terrestrial.eionet.europa.eu/CLC2000 for more information. The actual accuracy is still an open question, as we are currently focusing more on recognition value for navigation and performance. The latter part requires sometimes heavy simplification of the vectorised classification results in order to limit the number of triangles per tile. IIRC the maximum triangle count for the Oshkosh scenery is somewhere around 18.000 for some single tile (the others are around 10.000, but most are lower). Accuracy is obviously also dependent on the resolution of the imagery, which is 57m/pixel (one of the infrared bands, IIRC) via 28.5m/pixel (most bands, inclunding the visible light ones) to 14.5m/pixel (the panchromatic band). The panchromatic band can be used to enhance the visible light bands to double the resolution, but in the area of airports I don't think the result would be satisfying. This also means that some smaller regions such as tiny lakes or thin rivers may not be recognised or present in the final dataset, either because their shores blend too much with the surrounding terrain in the imagery or because they are removed on simplification (or both). Some of this may be an issue of more sophisticated training, but we are also investigating into how we could improve the simplification, e.g. by introducing weights so that a border between evergreen and deciduous forest could be simplified stronger than e.g. a border between lake and non-lake areas. We presented an experiment in the Berlin area with this approach on LinuxTag and I was told (as I wasn't able to go there myself) that we got very positive feedback. The Berlin scenery can be found here: http://www.custom-scenery.org/Berlin-Scenery.329.0.html > What we could also improve on the scenery in a much more simple step > would be to improve the textures. The current textures are much too > contrast-poor. At least that's my impression when I compare what I see > from above and what I see in FlightGear. Unfortunately, I'm not good at > knowing in advance what will look good, I just see when it doesn't look > good. > > > The current texture set is a huge improvement over what we had before, > which was a huge improvement over what we had before that, etc. etc. but > yes, there is still plenty of room for additional improvements in the > textures. Also, we really need to figure out a mechanism to blend the > transition between textures so we don't have the hard edges we have now. I very much favour the use of generic textures over the draping of satellite or aerial imagery, as generic textures can still provide almost arbitrary detail without using much space on disk and in RAM. When flying and navigating the important part is that a road, a forest or a city border is in approximately the right position. It is not important whether a specific tree is at the actual position it is in in reality. Furthermore, satellite or aerial imagery is only available freely for some regions of the world (parts of the U.S., for example) or only in low resolution or both. Still, by using generic textures and automatic classification, we can make much better use e.g. of the freely available Landsat imagery, even though the original data is only of a low resolution. Cheers, Ralf ------------------------------------------------------------------------- This SF.net email is sponsored by DB2 Express Download DB2 Express C - the FREE version of DB2 express and take control of your XML. No limits. Just data. Click to get it now. http://sourceforge.net/powerbar/db2/ _______________________________________________ Flightgear-devel mailing list Flightgear-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/flightgear-devel
