Video based convolutional neural networks forecasting for rainfall forecasting

This study presents a new methodology for improving forecasts of current monthly, regional precipitation using video-based convolutional neural networks (CNNs). Using 13 administrative regions of Great Britain as a case study, three CNN architectures are trained for each region to forecast monthly rainfall totals given forecast mean sea-level pressure and 2-m air temperature videos from the MetOffice GloSEA5 model and a benchmark rainfall data. The forecasts generated by the CNN and the GloSEA5 precipitation forecasts are both compared directly against a benchmark rainfall dataset for each of the regions. Following this, the CNN models are combined with the GloSEA5 forecasts to generate a new ensemble for each region which is then compared with the benchmark rainfall. The results show that the trained CNNs produce errors similar to the GloSEA5 model with RMSEs of 63 mm (single frame), 44 mm (slow fusion), and 37 mm (early fusion) compared with the GloSEA5 error of 33 mm. Regional variability remained consistent throughout the compared models. However, the CNN models all outperform GloSEA5 in the prediction of extreme events. Furthermore, treating the forecasts as an ensemble results in errors of 32 mm (CNN ensemble) and 31 mm (post-processing ensemble), both of which improve on the independent GloSEA5 forecasts