Novel statistical downscaling emulator for precipitation projections using deep Convolutional Autoencoder over Northern Africa

This study employed Machine Learning (ML) technique known as Convolutional Autoencoder to build Statistical Downscaling Model (SDM) emulator. Eight General Circulation Models (GCMs) rainfall datasets were selected under the Representative Concentration Pathway (RCP4.5) emission scenario over Northern Africa. Historical rainfall simulation for the period 1951?2005 from 8 GCMs were applied to train/validate the SDM. To evaluate the SDM performance emulating latest Rossby Centre (RCA4) RCM, SDM results were investigated against RCM projection products (2006?2100). Continuous statistics were employed to examine the SDM performance. The SDM has exhibited positive correlation of 0.75 < R < 0.95 and low RMSE values ranging between 6.9 and 15.8 mm/month. Similarly, the bias ratio scored a low value ranging from -8.94 < bias <8.25. The SDM showed good performance in reproducing the temporal rainfall projections, whereas unsatisfactory simulation was recorded regarding the spatial rainfall projections. In conclusion, the SDM showed better performance reproducing the projections of the mean ensemble rather than the individual RCMs. For future work, the SDM could be employed to downscale the mean ensemble projections of different climate variables.