Lot of variations in some places where topography varies a lot

Example on 31CTH, we can observe such sin(LIA) variations (around pixel {8090,4840}) .

From one pixel to the next, the LIA angle may jump from 3° to 120°.

If we look closely at normals produced, we see the same kind of variations: .

If we change the normals computation method to do cross products between averaged sides, a smoothing effect appear, but doesn't really solve the issue.