Postprocessing of rainfall forecasts over East Africa

We test methods of postprocessing rainfall forecasts out to 7 days over East Africa.

Using the physical forecast models, IFS from ECMWF and GFS from NCEP, we apply several combinations of post-processing techniques to empirically correct the predicted rainfall towards IMERG blended satellite rainfall data. The techniques we apply include a generative adversarial neural network (GAN) model (Harris et al. 2022), isotonic distributional regression (EasyUQ, Walz et al. 2024), EMOS (Gneiting et al. 2005), linear regression, and the kernel density estimate. Other approaches are also considered, however for the purposes of practical operational forecasts, we mainly focus on computationally cheap methods. Because we are comparing against IMERG, our results compare favourably against fully empirical models, such as FuXi and Graphcast, that have been trained to predict ERA5.

Being computationally cheap, in an operational forecast cycle on a standard desktop computer, the GAN model can produce spatially correlated 1000 member ensembles from the input forecast data. from which we can display the distribution using a histogram. The other techniques also cheaply produce rainfall distributions. We compare the quality of these distributions using the CRPS, variogram score and reliability diagrams. Biases in the raw rainfall forecasts are most notably reduced over the large lakes, for example Lake Victoria, over mountains, Indian ocean, and other places of high rainfall. We find it difficult to reduce biases in dry regions and over the Congo rainforest.

Different empirical modelling methods are optimal for different physical phenomena, and there is no theory for the most accurate model without physical insight. We also observe that it is often possible to improve each of the models with various tweaks. Each of the tested approaches might improve in the future, and the question we are trying to answer is “what is the best practical model available today?”