Scalable Bias Correction Techniques Investigated with a Low-Dimensional Model

Observation bias is a pervasive and highly individualized issue for numerical weather prediction models. Consequently, bias correction of observations is crucial for accuracy of analyses, especially for remote measurements such as satellite radiances. In the absence of model bias, bias correction schemes that correct observations to a model background, such as variational bias correction techniques, would work nearly perfectly. However, when undiagnosed model bias is present, such bias correction schemes are subject to “bias reinforcement.” We aim to develop a strategy for correcting model and observation bias independently, to avoid bias reinforcement while considering computational limitations of operational data assimilation systems. We explore this new methodology using the Model III of Lorenz (2005), an idealized chaotic dynamical model which simulates the evolution of a scalar field based on advection, diffusion, and constant forcing at two distinct scales of motion. For our experiments we induce controllable amounts of observation and model bias to simulate known challenges in operational NWP. We then use analysis increment and innovation statistics collected over a training period to independently correct model bias, observation bias, or both in subsequent experiments. In this way, model and observation biases are decoupled, resulting in a bias correction scheme that does not reinforce model bias.