A customer want to create a Mosaic Dataset with 15 000 aerial images (JPEG 2000, 11 000 with 15 cm resolution, and 20 cm for the others). Each Image size is 14,3 Mo and the amount of data is about 205 Go.
He has a desktop machine with this capabilities :
Windows 10, 4 cores, 10GB RAM, Graphic adapter NVIDIA Quadro NVS 450
What do you recommend for the system requirements to create the Mosaic Dataset with this volume of aerial Imagery ?
Such a machine should be fine. Creating mosaic datasets is not very computationally intensive. My biggest concern here would be JP2. Many JP2 files are created poorly and focus on reducing data volume at cost of being very IO and CPU intensive to access. From a performance perspective it would be advantageous to convert first to MRF,TIF or COG (See Imagery Formats and performance).
There are various things that can be done to speed up the creation. When adding rasters to a mosaic dataset only the header of the imagery needs to be read. If performance is important this can be reduced. It is not clear from your description if the image are just pre-generated tiles or source imagery from a sensor. Note ff using the Table raster type (or Frame Table for non rectified imagery) then if you include parameters such as Rows.Cols of the image then even reading the header of the image can be skipped.
Also do not compute statistics. Not required unless you want to do color correction. If building footprints use by Geometry (not by radiometry) unless the data you have really does have irregular edges. (This might occur if the data was pre-orthorectified, which we do not recommend). What would take time would be creating the overviews. Do take care as to what pixel size you generate this. If too small then you will need to read alot of the data and create a large overview. If too large then performance later can be affected as the system would need to read too many files was users pan/zoom at medium resolutions.
I am not an expert in image processing using ArcGIS, but I suspect that if all one wants to do is view the imagery as a backdrop, the hardware you describe is likely OK. However, if any kind of bulk processing is needed, then one has to look at the amount of information that needs to be handled in RAM at one time.
You mention that the space taken by the data is a bit over 200Gb. That could be misleading. For instance, if each image is a 512 by 512 tile, 3 band image, 8 bit pixel depth, uncompressed space needed would be somewhere higher than 10 Tb. Again, I don't really know in detail how ArcGIS handles imagery, but this would be my main concern. Hence, Michael Volz's recommendation to try doing what you want to do on a small sample is the wiser path to follow.
I do have experience in simply displaying thousands of rasters that I imported into a mosaic dataset. (I'm talking about 8 Tb). But if I wanted to do any geoprocessing on the whole full resolution (say, raster to vector), my system froze, even though I have 32 Gb of RAM & 8 cores. I have had to split the data into manageable subsets. Of course, I could have been doing something wrong, but my work got done.
Thank you very much for all the answers,I have discover new tools and parameters to optimize performance. I will check with the custormer if he can process the images and reach his goal