How could I extract all peaks from a DEM (large mountainous area)? Peak means something like: a pixel which is sourrounded by other pixels with smaler elevation numbers in a area arround the "peak-pixel" of about 100m.
Use the grass module r.param.scale with the "param=feature" option. THis creates an output map with each pixel categorized as peak, ridge, channel, plane, etc. Peaks are given category 6 (ridges=5, etc)
A pure pixel based approach could be done with "r.mapcalc" ([..]) but will likely not lead to exciting results. Hence r.param.scale as suggested by Micha or this Addon: r.prominence which calculates the average difference between a central cell and its neighbors. It approximated the terrain 'ruggedness' by looking at average differences in elevation within a given neighborhood. See http://grasswiki.osgeo.org/wiki/AddOns/GRASS_6#r.prominence
I took the reverse approach to finding peaks, I found the peak pixels first, and am now trying to select those exceeding a certain prominence (as described on peakbagger.com, they appear to manually inspect topos to get their peaks).
A solution for finding the single pixel peaks is to use r.terraflow, followed by r.mapcalc looking for the minimum: Flow Accumulation == 1. As terraflow uses multiple flow directions this method eliminates ridge lines and gives you do with a single flow direction model like D8. This is slow (2 hours for a 15,000 x 15,000 cell DEM), and I am sure there is a more elegant way to do it, but it works.
I took a completely different approach. I exported as ASCII using How to produce a CSV file from .tif file with elevation data? and then found the peaks outside QGIS on a 5*5 template using my favourite computer language. In QGIS 2.18.2 this is select layer then Raster->Conversion->Translate and then set the output to ASCII XYZ from the drop-down menu when you specify the output file. I would have preferred Arc/Info ASCII grid (*.asc) format, but the fix for the gotcha described in the above link can't work the same way in 2.18.2.