A comparison of three model output statistics approaches for the bias correction of simulated soil moisture

Abstract

Errors in the physics schemes and parameters of a land surface model can lead to large errors/bias in simulated soil moisture. In addition, large bias in simulated soil moisture may be caused by soil lateral water flow that is not described in a land surface model. In the present study, three model output statistics (MOS) approaches (a linear regression based MOS, a rescaling based MOS and a cumulative distribution function matching based MOS) are proposed to correct these biases and errors at stations that have soil moisture observations. Results show that the biases/errors in simulated soil moisture were significantly corrected by the three MOS approaches. The best performance was obtained for the linear regression MOS model consisting of four separate linear regression MOS models for different seasons throughout the year (MOS-SN). However, for the linear regression MOS and the rescaling MOS, a few months' data may also be used for MOS correction when the same seasonal period is used for both model fitting and prediction. With respect to goodness of fit and performance, the linear regression MOS model is better than the rescaling MOS model, and the MOS-SN is slightly better than the nonlinear cumulative distribution function matching approach. For sites where soil moisture observations are much shorter, or more intermittent, than meteorological observations, the linear regression MOS approaches can be used to reconstruct missing historic soil moisture data. All MOS approaches can also be used to correct the errors/bias in forecast soil moisture from weather or climate models at sites with soil moisture observations.