An embedded and charset-unspecified text was scrubbed... Name: not available URL: <https://stat.ethz.ch/pipermail/r-sig-geo/attachments/20110317/e8972184/attachment.pl>
Lidar data classification
3 messages · Ervan Rutishauser, Georg Ruß, Tomislav Hengl
On 17/03/11 11:38:19, Ervan Rutishauser wrote:
Dear all,
I would like to perform a classification of a large lidar data points
acquired over a tropical forest plot network (500 km2).
I computed the mean canopy(tree) height by 5m x 5m cell (pixel) and 3 others
parameters (max height, height variation over 2 years, mean height in a 50m
x 50 m neighborhood)
I did most of this using the raster package.
Now, I have a kind of multispectral image with 4 layers and I would like to
perform a large-scale classification of the data based on these 4
parameters. My aim is to find out "homogeneous" canopy regions, for
instance: high canopy with low change over the 2 years, canopy gaps, areas
of recruitment, etc.
I tried to perfom a standard cluster analysis (hclust), but I could not
compute the dissimilarity matrix (dist) on such a big data set (80'000 rows
and 4 variables), even with a 64-bits PC.
The k-means (kmeans{}) classification works, but return me strange results
(4 main clusters north/south/east/west). I have seen that the biOps package
allowed to do isodata classification. However isodata{} required an image
and I don't know how to compute a 4-layer image (if possible).
Hi Ervan, it seems you're up to doing some clustering on spatial data (which is similar to classification here). In other words, you have a spatialPointsDataFrame, and each of the points has four variables' values attached (canopyheight, max height, ...). The spatial data points are uniformly (or on a grid) distributed in space, I guess. If the above is correct: I've developed something that can perform exploratory clustering on this type of data sets. Nearly the same type of data occur in precision agriculture and they want to find management zones (management zone delineation): homogeneous areas inside a field. I'm currently working on that task and it's part of my PhD thesis. If you want, you can have a look at this publication of mine at last year's precision agriculture conference: http://fuzzy.cs.uni-magdeburg.de/aigaion/index.php/publications/show/772 I've written all of this in R and maybe we can have a look at the clustering I've done. I think what you want is probably something that gives you a first look at the data and helps you in delineating your forest into zones, if that's what you want. Regards, Georg.
Research Assistant Otto-von-Guericke-Universit?t Magdeburg research at georgruss.de http://research.georgruss.de
Op 17-3-2011 11:38, Ervan Rutishauser schreef:
Dear all,
I would like to perform a classification of a large lidar data points
acquired over a tropical forest plot network (500 km2).
I computed the mean canopy(tree) height by 5m x 5m cell (pixel) and 3 others
parameters (max height, height variation over 2 years, mean height in a 50m
x 50 m neighborhood)
I did most of this using the raster package.
Now, I have a kind of multispectral image with 4 layers and I would like to
perform a large-scale classification of the data based on these 4
parameters. My aim is to find out "homogeneous" canopy regions, for
instance: high canopy with low change over the 2 years, canopy gaps, areas
of recruitment, etc.
I tried to perfom a standard cluster analysis (hclust), but I could not
compute the dissimilarity matrix (dist) on such a big data set (80'000 rows
and 4 variables), even with a 64-bits PC.
The k-means (kmeans{}) classification works, but return me strange results
(4 main clusters north/south/east/west). I have seen that the biOps package
allowed to do isodata classification. However isodata{} required an image
and I don't know how to compute a 4-layer image (if possible).
Does anyone have any suggestion? Shall I turn me to GIS software?
I am trying to do it with SAGA at the moment, but find it difficult too :)
Are you referring to [http://www.saga-gis.org/saga_modules_doc/grid_discretisation/index.html]? Which problems did you experience exactly? I have been processing large grids using SAGA for years now. I am not a computer scientists, but if I test the same operation in SAGA and R, I usually discover that (a) SAGA is faster, (b) there are usually much less problems with memory consumption etc (see e.g. sec 5.5.2 in [http://spatial-analyst.net/book/]). SAGA has a fairly simply module development system [http://www.saga-gis.org/saga_modules_doc/lectures_introduction/index.html] and there is also an API functionality [http://www.saga-gis.org/saga_api_doc/html/] so you can access SAGA grids and libraries from python. As far as I know, linking of R and SAGA also goes pretty smooth. The drawback T. Hengl http://www.wewur.wur.nl/popups/vcard.aspx?id=HENGL001
Thank you for any help. Best regards, Ervan