Knowledge of snow's spatial distribution in terms of snow water equivalent (SWE) is important for hydrological forecasting, but current SWE products commonly disagree on regional scales. Assimilating passive microwave observations into a forecast from a physically-based snow model has been suggested to reduce or remove this disagreement, in which case the snow model must produce properties that are relevant to radiative transfer in snow. Here, the SNOWCAN model produces profiles of grain size for comparison with field measurements using contact probe spectroscopy and the impact of considering grain shape or conglomeration type (chain or cluster) is estimated. Prediction error in near-infrared optical grain size is estimated to be ±0.094 mm for all grains, or a possible best-case of ±0.083 mm if grain shape is included. The Helsinki University of Technology microwave radiative transfer model is used with the Cold Land Processes Experiment field data to make a preliminary estimate of the associated errors in simulated microwave brightness temperature difference, which is commonly used in SWE products such as Globsnow. Grain size error is associated with a ±5.1 K error and including grain shape, at best, reduces this error to ±4.5 K. Increasing stratigraphic detail by simulating more layers is an alternative method to reduce error, and layering error is found to increase linearly with snow depth. A single-layer simulation of 100 cm depth is associated with a ±8.7 K error relative to a pack described at the measurement resolution, whereas a 2-layer model is associated with a ±3.9 K error. Further work is required to determine the impact of grain: shape in the microwave regime, rather than the near infrared, but these results suggest that increased stratigraphic detail is a higher priority than including grain shape in order to improve the assimilation of passive microwave observations.

Le modèle de neige est souvent dépendant du modèle hydrologique avec lequel il est couplé, ce qui peut favoriser la représentation du débit au détriment de celle de la neige. L'objectif est de rendre le calage du modèle de neige plus indépendant de celui du modèle hydrologique en restant facilement utilisable en opérationnel. Dans cette optique, un modèle contraint sur des données d'observations de la neige permettrait d'améliorer d'une part la robustesse des paramètres du modèle de neige et d'autre part la simulation de l'état du manteau neigeux. Dans la première partie de cette thèse, nous avons étudié et modifié le modèle degrés-jour semi-distribué CemaNeige afin qu'il puisse simuler de manière plus réaliste la variable de surface d'enneigement du bassin versant. Cette modification, couplée au calage du modèle sur des données de surface enneigée et sur le débit, a permis d'améliorer la simulation de l'enneigement par le modèle sans détériorer significativement les performances en débits. Nous alors ensuite débuté le développement d'un nouveau modèle de neige à l'échelle ponctuelle. Celui-ci se compose d'un modèle de rayonnements, simulant les rayonnements incidents à partir de données d'amplitude de températures journalières, et d'un modèle de manteau neigeux. Le modèle de manteau neigeux résout les équations de la chaleur au sein du manteau neigeux à l'aide d'une représentation spectrale du profil de température. Cette représentation permet de simuler les profils et gradients de températures en utilisant moins de variables d'état qu'une discrétisation verticale par couches. Pour mieux prendre en compte les mesures ponctuelles de neige, ce modèle devra être distribué. / Snow models are often dependent on the hydrological model they are coupled with, which can promote higher performance on runoff simulation at the expense of snow state simulations performances. The objective of this thesis is to make the calibration of the snow model more independent from the calibration of the hydrological model, while remaining easily usable for runoff forecasting. Calibrating snow model on observed snow data would on one hand improve the robustness of the snow model parameters and on the other hand improve the snowpack modelling. In the first part of this manuscript, we modified the semi-distributed CemaNeige degree-day model so that it can explicitly simulate the watershed snow cover area. This modification coupled with the calibration of the model on snow cover area data and on river runoff data significantly improved the simulation of the snow cover area by the model without significantly deteriorating the runoff performances. Then we started the development of a new point scale snow model. It is based on a radiation model, which simulates incoming radiations from daily temperature range data, and a snowpack model. The snowpack model solves the heat equations within the snowpack by using a spectral representation of the temperature profile. This representation simulates the temperature profile and gradients using fewer state variables than a vertical discretization of the snowpack. In order to be able to use point scale snow observations in the model, it should be distributed on the watershed.