DeepRain: a separable residual convolutional neural algorithm with squeeze-excitation blocks for rainfall nowcasting

Precipitation nowcasting is critical for mitigating the natural disasters caused by severe weather events. State-of-the-art operational nowcasting methods are radar extrapolation techniques that calculate the motion field from sequential radar images and advect the precipitation field into the future. However, these methods assume the motion field's invariance, and prediction is based solely on recent observations, rather than historical radar sequences. To overcome these limitations, deep learning methods such as convolutional neural networks have recently been applied in radar rainfall nowcasting. Although, the promising progress of using deep learning techniques in rainfall nowcasting, these methods face some challenges. These challenges include producing blurry predictions, inaccurate forecasting of high rainfall intensities and degradation of the prediction accuracy with rising lead times. Therefore, the aim of this study is to develop a more performant deep-learning model capable of overcoming these challenges and preventing information loss in order to produce more accurate nowcasts. DeepRain is a convolutional neural network that uses a spatial and channel Squeeze & Excitation Block after each convolutional layer, local importance-based pooling, and residual connections to improve model performance. The algorithm is trained and validated using the UK Met Office's radar rainfall mosaic, which is produced by the UK Met Office Nimrod system. Using verification metrics, the model's predictive skill is compared to another deep learning model and a range of extrapolation methods.

Keywords: deep learning; rainfall nowcasting; radar; convolutional neural networks; Squeeze-and-Excitation