Assessing parsimonious hydrological model structures with distributed adjoint-based calibration in SMASH Python-Fortran platform on large sample of French catchments and flash floods

<p><span>This</span> <span>contribution</span> <span>presents</span> <span>improvements</span> <span>of</span> <span>conceptual</span> <span>models</span> <span>in </span><span>SMASH (Spatially distributed Modelling and ASsimilation for Hydrology) plat</span><span>form,</span> <span>underlying</span> <span>the</span> <span>French</span> <span>national</span> <span>flash</span> <span>flood</span> <span>forecasting</span> <span>system</span> <span>Vigicrues </span><span>Flash</span> <span>[1],</span> <span>based</span> <span>on:</span> <span>(i)</span> <span>the</span> <span>3-parameters</span> <span>model</span> <span>formulation</span> <span>and</span> <span>variational </span><span>data assimilation algorithm of [2] that showed promising results (i) hypothesis </span><span>testing on a large sample of catchments and flash floods; (ii) comparison of </span><span>the SMASH model performances in uniform and distributed calibration to GR </span><span>models;</span> <span>(iii)</span> <span>a</span> <span>new</span> <span>wrapped</span> <span>Python</span> <span>interface</span> <span>automatically</span> <span>generated</span> <span>by</span> <span>the </span><span>f90wrap library [3]. </span><span>Multiple tests have allowed us to converge on two parsimonious distributed </span><span>model structures that have comparable performances to the GR models in spa</span><span>tially uniform calibration.</span> <span>These two structures, mainly based on GR operators </span><span>at</span> <span>the</span> <span>pixel</span> <span>scale,</span> <span>differ</span> <span>in</span> <span>the</span> <span>production</span> <span>operator,</span> <span>with</span> <span>the</span> <span>6-parameters</span> <span>structure being GR production and the 7-parameters structure being VIC pro</span><span>duction.</span> <span>Furthermore, the use of distributed calibration applied to these formu</span><span>lations</span> <span>via</span> <span>adjoint</span> <span>model</span> <span>resolution</span> <span>shows</span> <span>significantly</span> <span>better</span> <span>calibration</span> <span>per</span><span>formances</span> <span>without</span> <span>being</span> <span>less</span> <span>robust</span> <span>in</span> spatio-<span>temporal</span> <span>validation.</span> <span>Immediate</span> <span>work </span><span>deals</span> <span>with</span> <span>improving</span> <span>the</span> <span>regional</span> <span>calibration</span> <span>scheme</span> <span>by</span> <span>tayloring</span> <span>the</span> <span>global</span> <span>search</span> <span>of</span> <span>semi-distributed</span> <span>prior</span> <span>parameter</span> <span>sets,</span> <span>with</span> <span>multi-gauge</span> <span>constrains, </span><span>improving physiographic regularizations in the forward-inverse SMASH assimilation</span> <span>chain, using Python librairies. </span></p><p><strong><span>References</span></strong><br><span>[1]</span> <span>P. Javelle, et al.</span> <span>Flash flood warn</span><span>ings:</span> <span>Recent achievements in france with the national vigicrues flash system UNDRR GAR, 2019.<br>[2] M. Jay-Allemand, et al.. On the potential of variational calibration for a fully distributed hydrological model: application on a mediterranean catchment. HESS, 2020, https://doi.org/10.5194/hess-24-5519-2020<br>[3] J. R. Kermode. f90wrap: an automated tool for constructing deep python interfaces to modern fortran codes. 2020. https://doi.org/10.1088/1361-648X/ab82d2<br></span></p>