Rain over Africa. An application of quantile regression neural networks to retrieve precipitation from geostationary satellites
Examensarbete för masterexamen
Complex adaptive systems (MPCAS), MSc
With an increasing change in Earth’s climate, global precipitation assessments are essential to understand the hydrological cycle and prepare society for extreme weather events. As of today, many countries are equipped with networks of weather stations that individually measure precipitation over small areas. However, adequate spatial coverage is not feasible for large regions like the African continent. This problem can be overcome by utilising geostationary satellites. They scan large surfaces of the Earth with a short revisit period, making them suitable for global nowcasting. Unfortunately, direct measurements cannot be made as surface precipitation rates need to be derived from an indirect relation to measured infrared (IR) and/or visible radiation. The problem is addressed by applying a Quantile Regression Neural Network (QRNN) to predict the a posteriori distribution of precipitation rates over Africa. The prob abilistic retrievals yield not only estimated precipitation rate values but also their associated uncertainties. Surface precipitation rate estimates with a 3 km × 3 km spatial resolution are achieved for Africa every 15 minutes. The QRNN is imple mented with a convolutional neural network (CNN) trained on thermal IR data from the SEVIRI instrument onboard the geostationary Meteosat-11 satellite. Addition ally, the GPM DPR and GMI combined precipitation product retrieved from the low Earth orbit GPM Core Observatory satellite is used as reference data. Two set-ups are trained, first using all thermal IR channels of the SEVIRI instru ment, and secondly using only two (6.2 µm and 10.8 µm). The second set-up indi cates promising use of historical data, while the first set-up shows the overall best performance. Comparisons are made with IMERG, a leading global precipitation product. The posterior distributions retrieved by the QRNN are shown to be more revealing than the scalar-valued estimations from IMERG. Moreover, it is possible to increase the spatiotemporal resolution as well as reduce the latency of the re trievals. The relative simplicity of the methodology hints at potential improvements in global precipitation products available today.
quantile regression, neural networks, CNN, precipitation, SEVIRI, GPM, nowcasting, Africa.