Weather radars provide real-time measurements of precipitation at a high temporal and spatial resolution and over a large domain. A drawback, however, it that these measurements are indirect and require careful interpretation to yield relevant information about the mechanisms of precipitation. Radar observations are an invaluable asset for the numerical forecast of precipitation, both for data assimilation, parametrization of subscale phenomena and model verification. This thesis aims at investigating new uses for polarimetric radar data in numerical weather prediction. The first part of this work is devoted to the design of an algorithm able to automatically detect the location and extent of the melting layer of precipitation , an important feature of stratiform precipitation, from vertical radar scans. This algorithm is then used to provide a detailed characterization of the melting layer, in several climatological regions, providing thus relevant information for the parameterization of melting processes and the evaluation of simulated freezing level heights. The second part of this work uses a multi-scale approach based on the multifractal framework to evaluate precipitation fields simulated by the COSMO weather model with radar observations. A climatological analysis is first conducted to relate multifractal parameters to physical descriptors of precipitation. A short-term analysis, that focuses on three precipitation events over Switzerland, is then performed. The results indicate that the COSMO simulations exhibit spatial scaling breaks that are not present in the radar data. It is also shown that a more advanced microphysics parameterization generates larger extreme values, and more discontinuous precipitation fields, which agree better with radar observations. The last part of this thesis describes a new forward polarimetric radar operator, able to simulate realistic radar variables from outputs of the COSMO model, taking into account most physical aspects of beam propagation and scattering. An efficient numerical scheme is proposed to estimate the full Doppler spectrum, a type of measurement often performed by research radars, which provides rich information about the particle velocities and turbulence. The operator is evaluated with large datasets from various ground and spaceborne radars. This evaluation shows that the operator is able to simulate accurate Doppler variables and realistic distributions of polarimetric variables in the liquid phase. In the solid phase, the simulated reflectivities agree relatively well with radar observations, but the polarimetric variables tend to be underestimated. A detailed sensitivity analysis of the radar operator reveals that, in the liquid phase, the simulated radar variables depend very much on the hypothesis about drop geometry and drop size distributions. In the solid phase, the potential of more advanced scattering techniques is investigated, revealing that these methods could help to resolve the strong underestimation of polarimetric variables in snow and graupel.