CDO

What about conservative remapping? see remapcon or gencon

You can also try gridboxmean, use

cdo gridboxmean,2,2 ifile ofile

Thanks for your help. I was thinking about remapcon
I am trying to understand what remapcon does exactly reading the doc file, but it is not that easy. Do you know by chance a link with illustrated example of this function?
In my case, for variables like runoff I need to go from a curvilinear (5*5km) grid to a lonlat (0.5*0.5degree) one.

conservative remapping computes for each target cell the weighted sum (by contributing area) of all contributing source cells.

I think this is indeed the good function, however I got an error message when trying:

cdo remapcon,target_grid.txt input.nc output.nc

cdo remapcon (Abort): grid1 corner missing!

It seems to be a prb my one of the grid (surely the 2nd one):

1. gridID 0 #
   gridtype = lonlat
   gridsize = 8584
   xname = lon
   xlongname = longitude
   xunits = degrees
   yname = lat
   ylongname = latitude
   yunits = degrees
   xsize = 116
   ysize = 74
   xfirst = -24.75
   xinc = 0.5
   yfirst = 35.25
   yinc = 0.5

input.nc grid

1. gridID 0 #
   gridtype = curvilinear
   gridsize = 550800
   xname = lon
   xlongname = longitude
   xunits = degrees_east
   yname = lat
   ylongname = latitude
   yunits = degrees_north
   xsize = 680
   ysize = 810
   xvals = -9.46942234 -9.41638756 -9.36334133 -9.31028366 -9.25721455 -9.20413494
   -9.15104294 -9.09793949 -9.04482555 -8.99170017 -8.93856335 -8.88541508
   -8.83225632 -8.77908516 -8.72590351 -8.67271137 -8.61950779 -8.56629276
   -8.51306629 -8.45983028 -8. etc............

The conservative remapping with remapcon needs the coordinates of all gridcell corners for the source and the target grid. For a regular lonlat grid these corners can be recomputed in CDO. But this is not possible for a curvilinear grid.