Probabilistic near real-time retrievals of Rain over Africa using deep learning

We introduce Rain over Africa (RoA), a public retrieval algorithm providing near real-time precipitation estimates over the entire African continent. The retrievals are based on Meteosat (geostationary) thermal infrared observations. Therefore, rain can be monitored constantly, just minutes after the dissemination of the input data. Despite this low latency, the accuracy of RoA is superior to estimates that require hours or more to obtain. Consequently, RoA is of particular interest where a rapid response is critical, such as for disaster preparedness.   The RoA retrievals employ a convolutional and quantile regression neural network. The latter allows for a detailed case-specific description of the retrieval uncertainty. Four years of data from the calibration satellite in the Global Precipitation Measurement (GPM) mission were used as labels for training and evaluation. With this deep learning setup, limitations in earlier near real-time retrievals for Africa were overcome. Moreover, the RoA network can be run on a regular workstation.   The effective resolution of RoA, at 50 km, renders its retrievals more timely and detailed than the well-established IMERG precipitation estimates. Additionally, the probabilistic nature of RoA enables addressing the inherent uncertainties of satellite precipitation retrievals by using probabilities of exceeding precipitation thresholds. Further assessment reveals largely similar diurnal cycles between RoA and IMERG, although IMERG shows some instability. Visual inspection of rain evolution patterns also indicates that RoA is more consistent. Finally, an annual mean analysis including CHIRPS estimates shows regional differences among the three, with no clear outlier behaviour for RoA.