We study the percolation transition in a longe-range correlated system
: a self-affine surface. For all relevant physical cases (i.e. positiv
e roughness exponents), it is found that the onset of percolation is g
overned by the largest wavelength of the height distribution, and thus
self-averaging breaks down. Self-averaging is recovered for negative
roughness exponents (i.e. power-law decay of the height pair correlati
on function) and, in this case, the critical exponents that characteri
ze the transition are explicitly dependent on the roughness exponent a
bove a threshold value. Below this threshold, the spatial correlations
are no longer relevant. The problem is analytically investigated for
a hierarchical network and by means of numerical simulations in two di
mensions. Finally, we discuss the application of those properties to m
ercury porosimetry in cracks.