GLnet:融合全局和局部信息的短临降雨预报网络

GLnet: Precipitation nowcasting network combining global and local information

基于雷达回波外推的定量降水预测具有广泛的应用前景。为了提高降水区域和强度的预测准确性,本文提出了一种新的基于Unet和Swin-Transformer的临近降水预报模型GLnet。该模型具有非对称双路特征提取结构,通过卷积和窗口自注意力机制分别提取雷达回波图片的局部和全局特征。同时在两类特征融合前引入了CBAM注意力机制和Non-local非局部注意力机制。本文在公开的荷兰降水地图数据集上分别采样出至少包含20%和和50%降水像素点的子集NL-20和NL-50,并利用结构相似性损失函数进行了实验。结果表明本文模型相比原始的Unet, MSE误差分别下降了14.4%和10.6%。

Quantitative precipitation prediction based on radar echo extrapolation has broad prospects.It’s important to get accurate nowcasting.To this end,we propose GLnet,an efficient neural networks-based on Unet and Swin-Transformer architecture equipped with two different attention modules CBAM and Non-local.The model has an asymmetric two-way feature extractor.In this way,the GLnet model extracts local and global features of radar echo images through convolution and windows self-attention mechanisms respectively.We create two datasets,NL-20 and NL-50,in Netherlands Precipitation Dataset by filtering the original precipitation dataset and choosing only the images with at least 20%and 50%of pixels containing any amount of rain respectively.We evaluate our approaches in NL-20 and NL50.The experimental results show that compared with the classical model Unet,the mean square error is reduced by 14.4%and 10.6%respectively.