一种多气象特征融合的时空降尺度模型

A Spatio-Temporal Downscaling Model Based onMulti-meteorological Feature Fusion

针对现有的深度学习气象时空降尺度方法较少、数据融合方式较为单一、气象图像纹理细节信息重建效果不佳等问题,提出了一种基于多特征融合的时空降尺度模型(Spatio-Temporal Downscaling Model Based on of Multi-Feature Fusion,STMFF)。设计一个动态数据融合模块(Data Fusion Module,DFM)对动态辅助数据进行融合,对特征时间插值模块进行改进,实现更加有效的静态数据融合;使用局部时间特征比较(Local Time Feature Comparison,LFC)模块来提高视频帧插值后每帧的图像质量,并采用基于多尺度特征提取的残差Swin Transformer模块(Residual Swin Transformer Module Based on Multi-Scale Feature Extraction,MF-RSTB)获得每帧图像的多尺度特征。基于法国雷达降水公开数据集的验证表明,STMFF最优结果较对比算法的MSE最大降低74.42%,最小降低4.28%,且在每帧图像的主观视觉效果上也更具优势。

To deal with the problems that the existing deep learning meteorological spatio-temporal downscaling methods are fewer,the data fusion method is relatively single,and the reconstruction effect of meteorological image texture details is not good,etc.,a Spatio-Temporal Downscaling Model Based on Multi-Feature Fusion(STMFF)is proposed.A dynamic Data Fusion Module(DFM)is designed to fuse dynamic auxiliary data,and the feature time interpolation module is improved to realize more effective static data fusion.The Local Time Feature Comparison(LFC)module is used to improve the image quality of each frame after video frame interpolation,and the Residual Swin Transformer Module Based on Multi-Scale Feature Extraction(MF-RSTB)is used to obtain the multi-scale features of each frame.Validation is made based on French radar precipitation open data set.Experimental results show that compared with the contrast algorithms,the maximum reduction of MSE for STMFF is 74.42%and the minimum reduction is 4.28%,and the STMFF also has more advantages in the subjective visual effect of each frame.