Memory limit problems in R / import of maps
On Tue, Apr 22, 2008 at 6:49 AM, Tomislav Hengl <hengl at science.uva.nl> wrote:
Dear list, I know that much has already been said about the memory limit problems. If there is any progress about this problem, we would be interested to hear. In our project, we are importing 24 maps/bands, each consists of 1,450,000 pixels. We further would like to glue all maps into a single data frame (e.g. 'ksc' class in adehabitat package; or 'SpatialGridDataFrame' in sp package), but this seems to be impossible. We tried to run this under windows (after following http://cran.r-project.org/bin/windows/base/rw-FAQ.html#There-seems-to-be-a-limit-on-the-memory-it-us es_0021 and setting the --max-mem-size) and under Linux Ubuntu, but still get the same error message (seems that there is no difference in use of memory under the two OS): "Error: Cannot allocate vector of size 11.1 Mb" The R workspace with 24 loaded grids is also really small (18 MB), but any further gluing and calculation is blocked due the vector size error message. For a comparison, in a GIS such as ArcGIS or SAGA/ILWIS (open source) we have no problems of loading and processing 3-4 times more grids. Should we simply give up on running spatial analysis using large grids (>10 million grids) in R?
Hi, What exactly were you hoping to do with such a massive data frame once you overcame the initial memory problems associated with loading the data? Any type of multivariate, classification, or inference testing would certainly require just as much memory to perform any analysis on the stack of grids. Not knowing what the purpose of this operation is (although I would guess something related to soil property or landscape modeling of some sort), it is hard to suggest a better approach. For grid that size I would use an algorithm that operates on strips or tiles. There are several great starting points in the GRASS source code. Doing all of the pre-processing, and possibly some aggregating to larger support size, in GRASS would allow you to test any R-centric operations on a coarser version of the original dataset. Cheers, Dylan