Adaptive Model Parameter Estimation Triggered by the Beneficial Observation Rate from Forecast Sensitivity to Observations

Takumi Honda , Yohei Sawada

In NWP, the assimilation of various observations contributes to improving forecast accuracy. The contribution of each observation can be estimated by existing methods. Empirically, it is well known that only a fraction of assimilated observations are diagnosed as beneficial, meaning that they improve forecast accuracy. Previous studies have indicated that the beneficial observation rate depends on model imperfections. Building on this sensitivity, we propose a new method that uses the beneficial observation rate to trigger adaptive model parameter estimation for mitigating model errors and bias. Specifically, the method activates model parameter estimation when the beneficial observation rate exceeds a prescribed threshold. Using the Lorenz96 40-variable system, we demonstrate that the new approach successfully detects model bias and improves analysis accuracy, even when the true model parameter varies in time. Furthermore, we find that the beneficial observation rate is useful for detecting model bias even when the model and observations have similar biases, in which case the time-averaged observation-minus-background does not provide a clear signal of the bias.