The 213th R-CCS Cafe - part 2

Because the weather is a chaotic system, a small perturbation in the initial condition will grow exponentially with time. Therefore, an accurate initial condition is important for a good forecast. The initial condition can be improved by using data assimilation, which combines the model and the observations. The observed variables are not always the same as the model variables. For example, in satellite data assimilation, the model variables include temperature and pressure, but the observed variable is the satellite radiance (brightness temperature). In this case, we cannot directly compare the model variables with the observed variable. To make the data assimilation work, a process called ‘observation operator’ is required to obtain the simulated observations from the model variables, so that we can compare the simulated observations and the real observations directly and then update the model. A model called Radiative Transfer for TOVS (RTTOV) is usually used as the observation operator for brightness temperature. In addition, because there is a bias between the simulated radiance by RTTOV and the observed radiance, a bias correction procedure is applied. In this study, we developed a new observation operator using machine learning.
We first run an experiment for one month to assimilate the conventional observations and the satellite brightness temperature. The output data is used to train the machine learning model. We then run three experiments for the same month the following year to test the performance of the machine learning model. It was shown that compared to the traditional method based on RTTOV and a bias correction, our new method is slightly worse. However, compared to only assimilating conventional observations, assimilating additional brightness temperature by using the observation operator based on machine learning improves the result. Therefore, our method works well. Finally, a separate bias correction procedure is not required in our method because the machine learning model has learned the bias during the training.