Recently, it has been shown that quadratic programming techniques with
a smoothing constraint controlled by a single parameter lambda provid
e a framework in which the reconstruction of a unit hydrogra ph from n
oisy excess rainfall and runoff data can be easily implemented. Howeve
r, a remaining problem in this approach is on the choice of an optimal
value lambda of the smoothing parameter lambda so that lambda is not
too small, with as a result a unit hydrograph overly sensitive to nois
e in the rainfall and runoff data, or not too large, with as a result
information loss in the estimated unit hydrograph. In this paper, the
use of cross-validation for computing a lambda is investigated and ill
ustrated with rainfall-runoff data published in the literature.