A new land data assimilation system is currently under development at the Meteorological Research Division (MRD) of Environment Canada (EC). Based on a simplified variational technique with a 24-h minimization period, the Canadian Land Data Assimilation System (CaLDAS) assimilates screen-level observations (air temperature and humidity) in order to specify soil moisture and surface temperature initial conditions for EC's numerical prediction systems. In CaLDAS, observations and first guess are combined in an optimal manner with weights that depend on the errors associated with both types of information. Thus the quality of soil moisture and surface temperature analyses crucially depends on the specification of these errors. The background errors are linked to uncertainties in land surface initial conditions, ancillary data, atmospheric forcing, and land surface modeling. The observation errors, on the other hand, are related to the measurement process (i.e., instruments), the measurements representativity, and the observation operator. This last element can be further partitioned into contributions from ancillary data, input parameters, and the transfer model itself. In this study, we present an ensemble approach to estimate background and observation errors, examine spatial and temporal variability of the relative weights for each type of errors, and determine their influence on the final soil moisture and surface temperature analyses.