Video-Based Convolutional Neural Networks for Rainfall Forecasting

Monthly variations in rainfall often lead to extreme events causing substantial
damage to society. This damage is often caused by droughts and floods. Current
sub-seasonal rainfall forecasts are based on large ensemble models using complex numerical predictions which require large amounts of computing power. Despite this, they often fail to capture the extreme events. In this study daily mean sea-level pressure (MSLP) and 2m air temperature (2AT) forecast images across the North Atlantic are used to produce regional, monthly rainfall forecasts for Great Britain. For each month 28 MSLP and 2AT images are derived from the MetOffice GloSEAS5 daily forecasts. The target rainfall is derived from the CEH-GEAR (Centre of Ecology and Hydrology Gridded Estimates of Areal Rainfall) and is used as the benchmark to be aimed for. Three types of convolutional neural networks (CNN) are trialled at combining the image sets into a regional rainfall forecast. These architectures are named slow-fusion, early-fusion and single frame. For each of the three architectures a CNN is developed independently for each region. The CNNs are then evaluated against derived monthly precipitation forecasts from the MetOffice’s GloSEAS5 model. The results show that all three CNN architectures outperform the derived
forecasts with the early-fusion and slow-fusion models performing the best, with root mean-squared errors of 11.2mm and 8.16mm respectively. The worst performing forecasts were given by the MetOffice derivations and the single frame CNN with RMSEs of 43.8mm and 33.3mm respectively. Errors across all models are highest in the western regions of Great Britain and lowest in the east. Similarly errors are also higher in the winter months (December, January and February). However, across these high-error months and regions the slow-fusion and early-fusion models consistently have errors less than 50% that of the MetOffice derivations.