I'm running some statistical comparisons between several DEMs (SRTM, ASTER, ALOS...) through GRASS-GIS and Python. Dealing with millions of cells is not very practical, so I thought I could compare elevation values from random points. How many points should I use? 5000? 10000?
Of course, the more, the better. But then I went searching for any literature on the subject and couldn't find anything with a good answer.
Can someone point me a good article on this?
The is no one answer as the number is samples is distribution dependent. In a homogeneous terrain a small number of random samples could adequately capture the distribution. However, in complex terrains one would need considerably more samples.
You could use the statistical moments (median, variance, percentiles) to evaluate if you have captured the variation but ideally a direct comparison of the distributions is a more robust way to evaluate if you have captured the statistical variation of the DEM in the sample. This can be done via a probability density function, available in most statistical software.
